Amiibo Theory - Generational Learning

[Faora Meridian]

Hi, I'm Faora Meridian. You might remember me from the Hey, Who WAS That? Player-Finder Thread, or random For Glory matches online.

I'm here today to posit a different theory regarding the nature of Amiibo learning patterns. Before I begin this experiment, I'd like to point out that I'm Australian. That means I still have a little over twenty-four hours until I pick up my copy of glorious, glorious Smash Bros for Wii U. My lamentations are vocal and persistent. It also means that my accent is inherently superior to yours. Sorry! I don't make the rules!

Onward! To learning!

What exactly is the Generational Learning theory? Simply put, I believe that it may be possible to create Amiibo of higher gameplay potential by exposing them to recursive instruction. Put MORE simply, I want to train one Amiibo, have that Amiibo train ANOTHER Amiibo, and so on. This is entirely a mental pursuit at present, but I'm honestly kind of interested in seeing if it has any merits and I would be willing to engage in the experiment if there would be interest from you fine folks.

This would require at minimum two Amiibo of the same type (for my purposes, I will likely use my main, Samus, as a benchmark for this), consistent rules (3-stock matches, Final Destination, no items, Fox Only) and for simplicity's sake a lack of item feeding. The Amiibo, as they grow, will have standardized movesets and no alterations to strength, defense or speed.

The process would go thusly: the Player (that's me!) trains the first Amiibo (henceforth referred to as Generation 1) until it reaches lv50. From what we already know of Amiibo, this Amiibo would learn my methods and how best to counter me. As I'm a Samus main against a Samus Amiibo, this results in a Samus-killing-Samus (SA-X, anyone?).

I then introduce Generation 1 to the second Amiibo (creatively named Generation 2). Generation 1 trains Generation 2 until Generation 2 has reached lv50. Even at this early stage, if Amiibo learn and adapt from their training, Generation 2 should be able to best Generation 1 on a consistent basis. Generation 1 has taught Generation 2 everything it needs to win.

I could then use a third Samus Amiibo (or, to save my poor, poor wallet, the first Amiibo with a full wipe back to lv1) to become Generation 3, learning from Generation 2. If the theory holds (and has held to this point, admittedly), then Generation 3 should be superior to Generation 2. And so on, and so forth.



If there is enough interest in the theory before the game launches, I'll grab myself another Samus Amiibo and engage in SCIENCE for our collective benefit. It should be interesting to see just what the limits of Amiibo skill end up being in the end. Thoughts? Comment! Offer support! Produce counter-theories! Collate data! Flamebait!

 

[Shiliski]

While this idea has merit and I would love to test it out once I get more amiibo (and they should be coming in the mail any day now), I feel it's necessary to point out that Generation 1 would continue learning even as it teaches Generation 2. The Amiibo don't stop learning, and Generation 1 would always have more experience than Generation 2.

That said, without having 2 amiibos to really test it on, that's just speculation. I look forward to seeing your results as well as running this kind of experiment myself.

 

[Faora Meridian]

I believe that the learning rate for Generation 1 compared to Generation 2 would be slowed. Generation 2 would be learning from Generation 1 in how to best counteract it, and it would be learning from a pre-defined method. Generation 1, by contrast, would be using the existing method effectively for a considerable period of time. By the time Generation 2 starts beating Generation 1, Generation 2 would, I believe, be the superior player. Generation 1 might continue to learn, but potentially at a slower rate.

The theory does assume an efficacy cap once lv50 is reached, or at best a little leeway. While I know Amiibo learn even after lv50, I would posit that the learning is slower and broader; the Amiibo learns how to react to new stimulus, rather than refinement of the old.

 

[elmike]

Your idea sounds great!. I would really like to know more about your findings.

 

[The Stoopid Unikorn]

It makes total sense.

Since whichever generation learns from the previous one, it means they can find ways to outplay the previous generation. The only problem is that constantly training your amiibo with the same character only allows it to perform well on ditto matches, as it barely trained against other characters. For example, a Lv 50 Xth generation Fox amiibo who's aware of Samus's projectiles would have a better chance against a Lvl 50 Xth generation Samus who mostly did dittos, since he could use the reflector and Samus wouldn't expect it.

 

[Faora Meridian]

I've just finished training the first Samus Amiibo (named THEBABY) to lv50. I stopped at this point to share the baseline for Generation 1, before this experiment goes further. But first, a note on my playstyle and the control elements as I'm the one instructing Generation 1 and the elements will remain stable throughout the experiment.

I play Samus semi-passively, spacing and controlling with missiles and Charge Beam (and to a lesser extent, zair). I'm pretty grapple-happy with my recoveries, but I don't do much grabbing. I prefer to keep my distance, but I tend toward hyper-aggressive play when forced to confront in close-quarters. While I normally would not be above using bomb jumps to aid in my recovery, I don't recall a single use of bombs in any capacity while training the Amiibo. This makes some of the following points even more interesting.

The control elements used are as follows: games were 3 Stock, 8 minute matches (though they never went to Sudden Death and rarely lasted longer than 4 minutes). Items were completely off. The stage was Pyrosphere's Omega form for thematics and, frankly, rockin' music (MP2 Multiplayer, anyone?). These elements, as noted above, will not be changed during the experiment.

The following are the points that really stood out to me while I was training the Generation 1 Amiibo. Bear in mind that this list is probably incomplete; I was too busy actually playing to take too many notes. You know how it is; you play Smash, you're meant to stop, then you keep playing... anyway! Notes!

- Made use of bombs to protect itself on the edge and in the air from lv1. This behaviour never changed despite it not stopping me and my never using it. Hard-coded?
- Observed me jump off the edge when Amiibo died and grapple back. Began attempting this behaviour at lv5. Failed regularly. Began succeeding more regularly at lv30. Still failing occasionally at lv50.
- Began range-spam and distancing emulation at lv9. Relatively efficient use of spacing and raged abilities at lv21.
- First time stock was taken from me was at lv20. This was due to messing a grapple recovery. The mistake was emulated twice and discarded, only to reappear as a consistant thorn in the Amiibo's side. Issue stems from attempted grapple recovery while tumbling; this results in f-air instead of z-air.
- First took a stock from me legitimately at lv28. My shame is known.
- Taunted me after a kill at lv32. May have been seeing things. Had been taunting it, but hadn't seen it emulate the behaviour before. Didn't see it again. Probably just my imagination.
- Right up until lv50, it could only manage to take one stock off me per round. I never went down to one stock and never lost a match.
- Main strategy by lv50 seemed to be ranged passive with extensive projectile spam. Grapple recoveries were preferred to the point of dying when they failed; this behaviour was never trained out by lv50. Very defensive play with extensive (and excessive) rolling and air-dodging. Very rare shield-grabs and grabs in general likely a result of learning from my playstyle.

Saving this information here for future notes. I'll head off now and begin observing the training of the Generation 2 Amiibo (named for old in-joke Lord Smaus. Bleh heh heh Smaus!). Results will be forthcoming!

 

[Faora Meridian]

So, there's a lot more notes from the raising of Gen 2's Samus to lv50 than there were for Gen 1. This is partly because I wasn't an active participant in the matches themselves, and so I was able to more readily take notes as well as notice details I'd missed before. In this post, we're going to see the notes I took on how Gen 1 trained Gen 2, Gen 1's adaptability to Gen 2, and then we're going to launch them into a lv50/lv50 5-set match with the same rules as we've seen so far.

Let's get started with the notes:

- First match, at lv2, Gen 2 managed a grab combo against Gen 1. This was never done before.
- An accidental footstool in the first match (Gen 1 jumping off Gen 2) was replicated twice more in that match in quick succession. It had not previously come up.
- When Gen 1 accidentally killed itself in the first match, Gen 2 TAUNTED IT. I didn't see things this time; I paused mid-taunt to grab my laptop and write this damn note.
- Early observation is that Gen 2 is much more short-hop happy than Gen 1. Will probably spam projectiles with it if it learns from Gen 1 further.
- I really like Pyrosphere's music.
- At lv8, Gen 2 is showing combo awareness and evasiveness. This is far in advance of Gen 1's training.
- Gen 2 is attempting combos I performed against Gen 1 that were not implemented in Gen 1's strategy. They are mildly successful, even at <lv10.
- Gen 2 seems to use bombs much more sparingly than Gen 1, but with greater efficiency than its opponent.
- Most kills scored against Gen 2 in the early training levels seem to be with Charge Beam and f-smash. Gen 1 has not yet attempted to capitalize on the bair and dair that I used effectively against it.
- Gen 1 plays a patient game when left at range. It builds Gen 2's percent slowly and seems to want to kill only at 170+.
- By game four, Gen 2 has adopted Gen 1's useless air-bomb usage. Why you do this, Gen 2? Was I not a good enough teacher?
- Pyrosphere has some pretty nice aesthetics. Beautiful design. Biased Smasher.
- By lv15, Gen 2 doesn't seem to shield as effectively as Gen 1 did at the same point. This could be my memory failing, or it could be that Gen 1 is teaching Gen 2 to rely on rolling instead of shielding.
- Gen 2 is showing at <lv20 a knowledge of which projectiles will overwhelm others. It successfully struck Gen 1 with a Charge Beam shot that would overwhelm Gen 1's missile attack, though it still falls prey to this from time to time. This is far earlier than Gen 1 learned the same thing. It is also grappling the ledge with more success at <lv20 than Gen 1 does at lv50.
- Gen 1 developed an overuse of aerial projectile spam. This was passed on to Gen 2 with much frustration on my part around the lv20 mark.
- Gen 1 also seems to have latched onto my accidental firing of a Charge Beam shot in the wrong direction. The Amiibo has learned a behaviour that is distinctly unhelpful. It doesn't seem to have passed onto Gen 2 by lv20. Simultaneously, my jump-off-the-edge-and-grapple-on-opponent-death patterns seem to have been cast aside by Gen 1 and have not been passed onto Gen 2.
- A match with Gen 2 around lv20 timed out with a Gen 1 victory. It very nearly took a stock off Gen 1 near the end of the match.
- Whenever I grabbed Gen 1, I threw without striking. Gen 1 seems to strike twice before throwing, but Gen 2 in mid-lv20s tends toward my quicker throws, only sometimes striking.
- At lv25, Gen 2 made use of a charged Smash. It failed to strike, but I will observe further.
- At lv25, Gen 2 made a successful kill against Gen 1. It then taunted Gen 1. Proof: Amiibo can learn to taunt on kills/deaths.
- This is also much, much earlier than Gen 1 started to take a stock off me. Current testing seems to back up the Generational Learning theory.
- I doubt I'll get bored of Pyrosphere.
- Gen 1 seems to still be learning, as is known from the footstool jumps earlier. In spite of that, it is not adapting strategy to what Gen 2 is doing.
- In a match where Gen 2 started at lv31, it took the first stock.
- In the same match, Gen 1 broke Gen 2's shield. It didn't know what to do while Gen 2 was stunned, and eventually used a nair. Inefficient. Honestly, painful to watch the punish go to waste.
- Amiibo paused, mid battle, to glare at one another for two seconds. Forget teaching computers to love; we have taught them to hate.
- Gen 2 has defeated Gen 1 in a battle for the first time. Stocks at time were 2 remaining for Gen 2 and 1 remaining for Gen 1. Gen 1 was at 55% and Gen 2 was at 127%. The battle started with Gen 2 at lv31 and ended with it at lv34.
- Remember, Gen 1 never was able to defeat me. This is the first instance of the experiment where the student has defeated the master.
- The very next match, Gen 2 scored a low% kill on Gen 1 for an early lead. Naturally, it taunted. You have learned well, my young apprentice.
- Another Gen 2 shield broken. This time, Gen 1 just ran back and forth until Gen 2 recovered. "Should I... should I get a medic, or should I beat the snot out of you?"
- Gen 2 seems to make more use of the shield than Gen 1 does. This could account for the shield-breaking that has happened at the mid-lv30 mark.
- Kill-taunts don't always happen. The Amiibo seems to judge whether a kill warrants it by some means.
- Gen 1 seems to have adapted to some of Gen 2's strategies after it lost. It won following a close match. I expect the next match to be a win for Gen 2.
- Need a break from training Amiibo? Poptarts make for a good time.
- At about lv40, Gen 1 2-stocked Gen 2. Perhaps adaptational learning continues at lv50 to a greater extent than I thought.
- From the late 30's to early 40's, Gen 2 is displaying considerably worse judgement during engagements. If I didn't know better, I'd wonder if it was letting Gen 1 win.
- Gen 2 seems to have abandoned grabbing for the most part, like me, as of lv40. Gen 1, by contrast, still effectively grabs/smacks around Gen 2.
- Gen 2 continues to display no interest in meteor smashes, or the beautiful, beautiful back-air.
- I swear, they spend half the match in Morph Ball form. Is it really that comfy?
- Game (lv40-lv43) did not go to Sudden Death as I expected. Instead, Gen 2 won with a % advantage. That's two victories against its teacher prior to hitting lv50. Knowing that the current settings aren't allowing for SD, I'll keep them this way as a control element.
- Just watched the two Amiibo roll laps across Pyrosphere. Great way to spend a day, this!
- Gen 2 failed a grapple recovery and saved itself, only to be punished by Gen 1 and fail the NEXT grapple recovery, dying in the process. Sigh.
- The Amiibo's trained passivity is borderline rage-inducing. They will roll away from a position where they can punish the opponent in order to maintain a safer position on the battlefield. No real aggression from Gen 1 or Gen 2.
- Gen 2 has, throughout the lv40's range, begun to strike with grabs. This is only at high %.
- I keep hearing about Amiibo who constantly make their perfect shields. Gen 1 and Gen 2 both seem to only pull it off roughly 50% of the time, if that. Poor training on my part? I struggle with perfect shields sometimes. Investigate further.
- The match that brought Gen 2 to lv50 was a 2-stock victory over Gen 1.



So, now that they're both lv50 and thus have equal bonuses that should neutralize any stat advantages, the only way to see which is truly superior is to pit them against one another. For this I settled on a Best-of-5 series, so as not to destroy my entire day with this portion of the experiment. All five games were run. So! What were the results?

Game 1: Victory for Gen 2 by Sudden Death (apparently the time when I thought it wasn't on, it was!) Of particular note was that neither Amiibo hit 100% damage within the first two minutes of the match.
Score: 0-1

Game 2: Victory for Gen 2 by 2 stocks (126%). Fun fact: Gen 2 managed to take Gen 1 to 130% before taking any return damage. Gen 1 also ran around like an idiot again when Gen 2's shield was broken and self-destructed at low% during its second stock. When Gen 2 also broke Gen 1's shield, Gen 2 had no idea what to do, either.
Score: 0-2

Game 3: Victory for Gen 1 by Sudden Death. More Gen 2 shieldbreaking, more Gen 1 confusion. Perhaps it should try punching or kicking when the opponent is idle?
Score: 1-2

Game 4: Victory for Gen 2 by Sudden Death explosions (note that as items are off, they had no idea what to do with them. Gen 1 caught the first bomb to drop, but held on until it killed him). The Shieldbreak Heartbreak saga continues. This time at least, Gen 1 was able to punish it with a F-Smash. So, it turned out, was Gen 2 after seeing this done to it. Gen 2 also paused for about five seconds in the middle of the match for no obvious reason. Gen 1 seemed as confused as me by this.
Score: 1-3

Game 5: Victory for Gen 1 by 1 Stock (71%). Another broken Gen 2 shield went unpunished, to my dismay. Equally dismaying is the ongoing lack of taunting. I know they do it now, so why have they stopped? So disappointing.

FINAL SCORE: 2-3. Generation 2 has defeated Generation 1.



I don't feel confident enough to declare anything after this best-of-5 series. A lot of situations could have been slightly different and affected the outcome, however I'm going to continue on as planned and have Gen 2 train Gen 3. The reasoning behind this is the efficacy with which Gen 2 was able to, at an early point, become challenging for Gen 1. It holds that Gen 3 should be able to learn and adapt as fast, or faster, than Gen 2 if Gen 2 is the tutor.

Stay tuned! I'm gonna take a break for a bit and get back to it later!

 

[Faora Meridian]

Generation 3's training is complete and, frankly, the results have been some of the most interesting so far. I'll let my notes speak for themselves, and then we'll move into my thoughts on what it means.

- Gen 3 started at lv1 with aggression. It's getting its backside kicked, but it was using melee and projectile moves more experimentally than any generation before it. Perhaps resetting the Amiibo leaves something behind? If so, resets aren't complete and the experiment is invalid from Generation 3 onward.
- Does the new Amiibo know that it's against another Amiibo? Is it learning faster because it borrows from the other Amiibo's existing data? Experimentation required.
- Gen 2 attempted and almost succeeded in a dair spike. I was so, SO hoping it would work, but Gen 3 hit the ground on the edge of the stage.
- Gen 3 has begun bomb-dropping at strange intervals like its predecessors, and has done so without any prompting from Gen 2. Most likely a hard-coded thing for Samus Amiibos?
- Gen 3 attempting at lv10 fun things like d-tilt and bair strikes. They're not working against Gen 2, but that they're being tried is a good sign.
- Gen 3 is learning to perfectly shield strikes at lv11. This is much earlier than before.
- At lv13, Gen 3 is far more willing to wade into melee range than Gen 2. It's actually landing some decent blows, and has even pulled off a grab without any prompting from Gen 2.
- It's also managing to achieve grapple recoveries now, before Gen 2 has shown it how.
- At lv15, Gen 3 has scored a kill against Gen 2. It taunted with no prompt to do so. This is the earliest kill in the experiment to date.
- Gen 3 is dodging and shielding much more efficiently at lv19 than Gen 2 or 1 did. It's also still comfortable wading into melee range, even though that is usually where Gen 2 takes it down.
- Another kill by Gen 3, this time at lv19. Another taunt, too. Cocky bugger.
- At lv21, Gen 3 displayed less interest in grapple recoveries. Instead, it likes to use Screw Attack up past the ledge to take out the edge-guarding Gen 2. This seems to hit Gen 2 roughly 30% of the time, be punished 50% of the time, and land safely on the ledge or on the edge 20% of the time.
- At lv22, Gen 3 has begun to punish grapple recoveries by dropping down to bair Gen 2 while attempting to recover. This is an effective way of killing Gen 2, despite Gen 2's higher skill level and advantages.
- Gen 3 also displayed an ability to chain throws of Gen 2 at low%. It racked up about 25% damage with a forward throw to forward throw.
- At lv24, Gen 3 has displayed Smash-charging. It has yet to succeed, but will become more common if past patterns hold.
- At mid-lv20s, Gen 3 seems to display the ability to reliably take one stock off Gen 2 with each match. This is impressive compared to past examples in the experiment.
- Gen 2 broke Gen 3's shield. After first whiffing a charged f-smash, it landed the second charged f-smash before Gen 3 recovered. Seems this is how long it takes to learn how to handle shield breakages. Progress!
- I could really go for some more poptarts right about now. They're great Smash food!
- At lv26, Gen 3 dodged two grab attempts from Gen 2 before countering with a grab of its own that worked. Gen 3 is now using grabs more effectively than Gen 2, and without using zair and grapple recoveries.
- As if to make me eat my words, Gen 3 has begun to grapple recover. Jerk.
- Gen 3 seems to be leveling faster than Gen 2. Is it because it's being exposed to more advanced strategies? More experimentation required.
- Gen 2 becoming more aggressive with melee strikes. Even using f-smash to punish rolling. Learning from Gen 3's example, perhaps?
- All generations exhibiting a fundamentally useless propensity for firing projectiles in the air while their opponent fires/charges from the ground. They miss roughly 45% of the time and are dodged roughly 50% of the time. Why do they hold onto this behaviour? Dropping the bombs has some strategic use to ward off the opponent. This doesn't even effectively control the airspace. Why hold onto it?
- Gen 2 killed Gen 3 with a bair. Smells like... victory. Also foot.
- In spite of learning how to punish grapple recoveries at much lower levels, the lv31 Gen 3 seems to have abandoned that in favour of using grapple recoveries itself.
- Seriously, am I ever going to get sick of Ridley's Theme in all of its glorious arrangements? Theory suggests not.
- Gen 2 continues to reinforce the awesomeness that is the bair. Gen 3 has yet to learn this skill.
- Gen 3 just hit lv35, and I'm reminded that by this point Gen 2 had managed to win a game against Gen 1. Has Gen 2 become more viable through extensive, adaptive play? More experimentation required.
- Gen 3 continues to avoid punishing Gen 2's grapple recoveries, while Gen 2 has begun punishing Gen 3's grapple recoveries.
- Gen 3 is a disgrace to its country and brings shame upon its family.
- Gen 2 is now consistently punishing a broken shield. Yet to see if Gen 3 has learned this skill.
- Gen 2 is now actively trying to pressure and break the shield. This is new behaviour.
- Gen 2 is attractive and aromatically pleasing to those it encounters.
- Gen 3 has reached over lv40 and has still yet to win. It could have won the game that brought it to lv40, but decided instead that grapple recoveries were good and killed itself.
- I'm not a stoner, I just really like poptarts.
- Gen 2's old patterns of, "What do I DO?!" when breaking the enemy shield resurfaced for what I can only assume are digital giggles.
- Neither Gen 2 nor Gen 3 seem to have any interest in zair except for grapple recoveries. Both Gen 1 and I were able to use them reasonably well. What changed? Investigation required.
- At lv43, I should not be surprised when Gen 3 takes a stock off Gen 2. Instead, I am. Troubling.
- Coming up on lv50 and it still hasn't won a match. Frustration is building. Compounded by a Gen 3 self-destruct. You just did that last game, Gen 3. Did you think it would inspire sympathy? Have the simulated mind games begun?
- A broken shield enticed Gen 3 to attack with a down-smash. Not the best option, but at least it wasn't running around like a headless chicken. Progress, I suppose.
- Gen 3, this is Adam. I authorize you to win now.
- ... did that work? No? Damn. What does it take, Gen 3?
- All the way to lv50 and never once was able to defeat Gen 2. For contrast, Gen 2 first beat Gen 1 in the early lv30s.


From this, we can make some guesses about Amiibo behaviour from one generation to the next. The initial thought process that I had suggested that Amiibo could only adapt to a certain point. This I figured was erroneous; if Gen 2 was able to defeat Gen 1, why could Gen 3 not learn to defeat Gen 2? The patterns weren't any different; Gen 3 had to observe and adapt. It failed to do this on the route to lv50.

There is the possibility, as in my notes, that a wiped Amiibo retains some of what it knew before. If this is the case, Gen 3 may have been tainted by some of the old Gen 1 tactics that Gen 2 was able to defeat. This possibility is impossible for me to confirm or rule out entirely.

Finally, there is the potential that Gen 2, having been trained by Gen 1 and then gone through additional battles with it, was already well-suited to fighting another Amiibo. Gen 3 was learning these Amiibo-fighting techniques from Gen 2, but Gen 2 continued to perfect them. Indeed, whenever Gen 3 seemed to have an advantage, it was fleeting. Gen 2 seemed more capable of adaptational learning than even the relative blank slate of Gen 3.

Still! There is a chance that Gen 3's redemption will come during the Best-of-Five round!

Game 1: Gen 3 took the first stock, which surprised the hell out of me. I was all geared up for obliteration, but it actually held pretty well. Part of the why might have been Gen 2 returning to its old ways, failing to punish the broken shield. Gen 3 ended up taking the match, with 2 stocks at 112% to Gen 2's 1 stock at 116% at the end of time.
Score: 0-1

Game 2: Everyone mixed up their strategies here this time. Gen 2 decided the way to go was aggressive grabbing. Gen 3, by contrast, figured that its best path to victory was to be shot in the face a lot. As a result, Gen 2 won the match by 2 stocks. It wasn't even close, really.
Score: 1-1

Game 3: Gen 3 managed to get a great shield break on Gen 2 this match. Then, in the biggest display of bad manner behaviour from an AI in the history of gaming, Gen 3 proceeded to f-smash the air JUST in front of Gen 2's face until they recovered. That Gen 3 went on to win the match in Sudden Death with a Charge Beam shot just proves that it knew it could be a jerk and get away with it.
Score: 1-2

Game 4: I wish I had something funny to say about this game, but it was full of bad play from both Amiibo. I felt bad for both of them and wished them a speedy, fiery demise. In the end, Gen 2 took control of the match and won 2 stocks to 1 at time.
Score: 2-2

Game 5: Much like the previous game, there wasn't much to find I appreciated. Near the end, Gen 3 actually landed quite a few good blows to put it in a commanding position though, and while I'd like to say that it won with a spectacular d-air spike, it didn't. Instead, it won with a Charge Beam shot.

FINAL SCORE: 2-3. Generation 3 has defeated Generation 2.



And I did NOT see that coming! This is insanity. Despite all of the crap that happened in the training, Gen 3 wound up taking it. While, again, a sample size of five matches is hardly acceptable, the standard of its play was indeed slightly improved from Gen 2 by the end. While I'm unsure of what the future holds for the experiment, it's plain that it can continue for now.

Before I wipe Gen 2 and start again with Gen 4, I will be doing something that, by all arguments, will contaminate the experiment. I will be taking Gen 2 and fighting it myself in another Best-of-5 to see if I am able to best it. If you recall, Gen 1 was unable, at any point, to defeat me by the experiment's ruleset. Indeed, it could only ever take one stock off me. If Gen 2 is capable of giving me a greater challenge, that will speak much to the experiment's worth.

That will have to come later, as there's a storm rolling in. The experiment will return with all haste, though, so leave your thoughts, comments and suggestions!

 

[Dark Dr. Pink-Gold Pit Jr]

Wow, this is really cool! But that means...


...rip wallet.

 

[Trekkerjoe]

Awesome work. I'm going to follow this closely.

 

[kyxsune]

@ Faora Meridian

So when the AI take over, they will see a record of us pitting their primordial ancestors against themselves in combat...for research. Keep it up, this is fascinating.

Send me a PM if you need/want financing for 4th gen and up.

 

[Faora Meridian]

Wow, this is really cool! But that means...


...rip wallet.

Not so! For as I'm doing, I'm going to wipe Gen 1 back to a clean slate and use it for Gen 3. You only need two Amiibo of the same variety to pull this off!

Awesome work. I'm going to follow this closely.

Thanks; glad you like it! I'm about to get started on Gen 3 at the moment. Should have some results in a few hours!

@ Faora Meridian

So when the AI take over, they will see a record of us pitting their primordial ancestors against themselves in combat...for research. Keep it up, this is fascinating.

Send me a PM if you need/want financing for 4th gen and up.

It's always good to chart the genesis of our inevitable machine overlord's hatred for humanity's cruelty, I find!

As much as I appreciate the offer, it's unnecessary. As noted above, I can simply wipe the previous generations when they are no longer needed, and move forward like that. Thanks very much for the kind thought, though; it's very appreciated!

 

[The Stoopid Unikorn]

So far, I'm really excited, but I'd like to see what happen to a Gen 2 or later amiibo if any previous generation (or the player himself) started giving quick training sessions to a Gen X amiibo

 

[Squadaloo]

Very interesting study you have going here, can't wait to see more.

Also, this bit made me laugh quite a bit:

- Another Gen 2 shield broken. This time, Gen 1 just ran back and forth until Gen 2 recovered. "Should I... should I get a medic, or should I beat the snot out of you?"

 

[Faora Meridian]

So far, I'm really excited, but I'd like to see what happen to a Gen 2 or later amiibo if any previous generation (or the player himself) started giving quick training sessions to a Gen X amiibo

I would need more Samus Amiibo if I was to do that and, frankly, that's just a waste of money. I'll probably hold onto Gen 2 for a bit after this and see what happens to it if I train against it for a bit personally. I should be able to teach it a few things. Primarily, I'm tempted to go at it entirely at melee range and see if it picks up on that. More aggressive play would be infinitely more effective.

Very interesting study you have going here, can't wait to see more.

Also, this bit made me laugh quite a bit:

The results from the 1-50 training for Generation 3 are up above now, so read and enjoy! There is more of my humour there to entertain and delight, as well!

 

[erico9001]

Weird. I was expecting Gen 3 to beat Gen 2 and was hypothesizing the real issue would be when gen 3 faced you.

Could you face the amiibos against yourself and tell me how they fair? I want to know if the skill to beat the original person who trained them is retained or lost with each generation.

 

[The Stoopid Unikorn]

Generation 3's training is complete and, frankly, the results have been some of the most interesting so far. I'll let my notes speak for themselves, and then we'll move into my thoughts on what it means.

- Gen 3 started at lv1 with aggression. It's getting its backside kicked, but it was using melee and projectile moves more experimentally than any generation before it. Perhaps resetting the Amiibo leaves something behind? If so, resets aren't complete and the experiment is invalid from Generation 3 onward.
- Does the new Amiibo know that it's against another Amiibo? Is it learning faster because it borrows from the other Amiibo's existing data? Experimentation required.
- Gen 2 attempted and almost succeeded in a dair spike. I was so, SO hoping it would work, but Gen 3 hit the ground on the edge of the stage.
- Gen 3 has begun bomb-dropping at strange intervals like its predecessors, and has done so without any prompting from Gen 2. Most likely a hard-coded thing for Samus Amiibos?
- Gen 3 attempting at lv10 fun things like d-tilt and bair strikes. They're not working against Gen 2, but that they're being tried is a good sign.
- Gen 3 is learning to perfectly shield strikes at lv11. This is much earlier than before.
- At lv13, Gen 3 is far more willing to wade into melee range than Gen 2. It's actually landing some decent blows, and has even pulled off a grab without any prompting from Gen 2.
- It's also managing to achieve grapple recoveries now, before Gen 2 has shown it how.
- At lv15, Gen 3 has scored a kill against Gen 2. It taunted with no prompt to do so. This is the earliest kill in the experiment to date.
- Gen 3 is dodging and shielding much more efficiently at lv19 than Gen 2 or 1 did. It's also still comfortable wading into melee range, even though that is usually where Gen 2 takes it down.
- Another kill by Gen 3, this time at lv19. Another taunt, too. Cocky bugger.
- At lv21, Gen 3 displayed less interest in grapple recoveries. Instead, it likes to use Screw Attack up past the ledge to take out the edge-guarding Gen 2. This seems to hit Gen 2 roughly 30% of the time, be punished 50% of the time, and land safely on the ledge or on the edge 20% of the time.
- At lv22, Gen 3 has begun to punish grapple recoveries by dropping down to bair Gen 2 while attempting to recover. This is an effective way of killing Gen 2, despite Gen 2's higher skill level and advantages.
- Gen 3 also displayed an ability to chain throws of Gen 2 at low%. It racked up about 25% damage with a forward throw to forward throw.
- At lv24, Gen 3 has displayed Smash-charging. It has yet to succeed, but will become more common if past patterns hold.
- At mid-lv20s, Gen 3 seems to display the ability to reliably take one stock off Gen 2 with each match. This is impressive compared to past examples in the experiment.
- Gen 2 broke Gen 3's shield. After first whiffing a charged f-smash, it landed the second charged f-smash before Gen 3 recovered. Seems this is how long it takes to learn how to handle shield breakages. Progress!
- I could really go for some more poptarts right about now. They're great Smash food!
- At lv26, Gen 3 dodged two grab attempts from Gen 2 before countering with a grab of its own that worked. Gen 3 is now using grabs more effectively than Gen 2, and without using zair and grapple recoveries.
- As if to make me eat my words, Gen 3 has begun to grapple recover. Jerk.
- Gen 3 seems to be leveling faster than Gen 2. Is it because it's being exposed to more advanced strategies? More experimentation required.
- Gen 2 becoming more aggressive with melee strikes. Even using f-smash to punish rolling. Learning from Gen 3's example, perhaps?
- All generations exhibiting a fundamentally useless propensity for firing projectiles in the air while their opponent fires/charges from the ground. They miss roughly 45% of the time and are dodged roughly 50% of the time. Why do they hold onto this behaviour? Dropping the bombs has some strategic use to ward off the opponent. This doesn't even effectively control the airspace. Why hold onto it?
- Gen 2 killed Gen 3 with a bair. Smells like... victory. Also foot.
- In spite of learning how to punish grapple recoveries at much lower levels, the lv31 Gen 3 seems to have abandoned that in favour of using grapple recoveries itself.
- Seriously, am I ever going to get sick of Ridley's Theme in all of its glorious arrangements? Theory suggests not.
- Gen 2 continues to reinforce the awesomeness that is the bair. Gen 3 has yet to learn this skill.
- Gen 3 just hit lv35, and I'm reminded that by this point Gen 2 had managed to win a game against Gen 1. Has Gen 2 become more viable through extensive, adaptive play? More experimentation required.
- Gen 3 continues to avoid punishing Gen 2's grapple recoveries, while Gen 2 has begun punishing Gen 3's grapple recoveries.
- Gen 3 is a disgrace to its country and brings shame upon its family.
- Gen 2 is now consistently punishing a broken shield. Yet to see if Gen 3 has learned this skill.
- Gen 2 is now actively trying to pressure and break the shield. This is new behaviour.
- Gen 2 is attractive and aromatically pleasing to those it encounters.
- Gen 3 has reached over lv40 and has still yet to win. It could have won the game that brought it to lv40, but decided instead that grapple recoveries were good and killed itself.
- I'm not a stoner, I just really like poptarts.
- Gen 2's old patterns of, "What do I DO?!" when breaking the enemy shield resurfaced for what I can only assume are digital giggles.
- Neither Gen 2 nor Gen 3 seem to have any interest in zair except for grapple recoveries. Both Gen 1 and I were able to use them reasonably well. What changed? Investigation required.
- At lv43, I should not be surprised when Gen 3 takes a stock off Gen 2. Instead, I am. Troubling.
- Coming up on lv50 and it still hasn't won a match. Frustration is building. Compounded by a Gen 3 self-destruct. You just did that last game, Gen 3. Did you think it would inspire sympathy? Have the simulated mind games begun?
- A broken shield enticed Gen 3 to attack with a down-smash. Not the best option, but at least it wasn't running around like a headless chicken. Progress, I suppose.
- Gen 3, this is Adam. I authorize you to win now.
- ... did that work? No? Damn. What does it take, Gen 3?
- All the way to lv50 and never once was able to defeat Gen 2. For contrast, Gen 2 first beat Gen 1 in the early lv30s.


From this, we can make some guesses about Amiibo behaviour from one generation to the next. The initial thought process that I had suggested that Amiibo could only adapt to a certain point. This I figured was erroneous; if Gen 2 was able to defeat Gen 1, why could Gen 3 not learn to defeat Gen 2? The patterns weren't any different; Gen 3 had to observe and adapt. It failed to do this on the route to lv50.

There is the possibility, as in my notes, that a wiped Amiibo retains some of what it knew before. If this is the case, Gen 3 may have been tainted by some of the old Gen 1 tactics that Gen 2 was able to defeat. This possibility is impossible for me to confirm or rule out entirely.

Finally, there is the potential that Gen 2, having been trained by Gen 1 and then gone through additional battles with it, was already well-suited to fighting another Amiibo. Gen 3 was learning these Amiibo-fighting techniques from Gen 2, but Gen 2 continued to perfect them. Indeed, whenever Gen 3 seemed to have an advantage, it was fleeting. Gen 2 seemed more capable of adaptational learning than even the relative blank slate of Gen 3.

Still! There is a chance that Gen 3's redemption will come during the Best-of-Five round!

-- This space for future use. Don't hold your breath! Oxygen is good. --

My theory about why Gen 2 seems better: unlike Gen 1, Gen 2 didn't teach everything it knows to Gen 3, so it could have an upper hand (or should I say foot?) against Gen 3. So it could be smarter in terms of teaching things to avoid getting any "student beating master" situation

 

[Faora Meridian]

Weird. I was expecting Gen 3 to beat Gen 2 and was hypothesizing the real issue would be when gen 3 faced you.

Could you face the amiibos against yourself and tell me how they fair? I want to know if the skill to beat the original person who trained them is retained or lost with each generation.

Facing the Amiibo myself would contaminate the experiment going forward. I'm willing to do it with Gen 2's Amiibo before I wipe it to become Generation 4, but not Gen 3 just yet. Also, Gen 1 was never able to defeat me. It could only ever take one stock. Yay me!

My theory about why Gen 2 seems better: unlike Gen 1, Gen 2 didn't teach everything it knows to Gen 3, so it could have an upper hand (or should I say foot?) against Gen 3. So it could be smarter in terms of teaching things to avoid getting any "student beating master" situation

While this is an interesting theory, I'm not sure that the Amiibo have sophisticated enough programming to withhold techniques from other Amiibo learning from them. This would also go against the fundamental nature of the Amiibo: they're designed to learn from the play they are exposed to. Having one learn not to use advanced techniques would not only result in a weaker fighter, but it would also violate that fundamental nature.

They're still just Amiibo. I'm not creating Skynet, here*!

* Faora Meridian takes no responsibility for the advent of our machine overlords.

 

[erico9001]

Facing the Amiibo myself would contaminate the experiment going forward. I'm willing to do it with Gen 2's Amiibo before I wipe it to become Generation 4, but not Gen 3 just yet. Also, Gen 1 was never able to defeat me. It could only ever take one stock.

Well that's a bummer, but good to hear that you can do it in the future. Although gen 1 did not beat you fully, if one generation does better than another that would be enough to answer the question.

 

[Faora Meridian]

So. The storm passed. What did not, however, was my curiosity. As the lightning raged outside, I had a thought that led me to a very unsatisfying observation. I can show you all, each and every one with Smash, what my Amiibo Best-of-5 bouts were like. I can do this very simply.

Step 1: Put two lv9 Samus CPUs against each other.
Step 2: Set stage and match type to the same as the experiment's.
Step 3: Imagine a lot more dodging and rolling.

I realized while offline that the experiment lacked a baseline to compare it to. Creating a specialized Samus-killing Amiibo is one thing, but the nearest benchmark that's not me is a lv9 CPU opponent. By pitting two CPUs against one another, I could see how lower-level play was meant to ensue.

Unfortunately, I was disappointed. It was almost identical to the Amiibo bouts. There were notable differences, of course; the Amiibo were shielding and dodging better, for one. They also were a lot faster in their movements and more precise in certain ways. They were definitely superior in their matches to the lv9 CPU ditto.

What they were not, however, was different enough to impress me. In particular, note the moveset shifts I made mention of during the experiment. The Amiibo would often settle into projectile spam of Charge Beam and missiles, would jump to launch in the air, would drop bombs to (apparently) control space, and would sometimes get a few throws in. They'd grapple-recover, occasionally make a smash or tilt, and were loathe to use the bair or dair spike.

Observing a lv9 CPU ditto matchup shows the exact same moveset preferences.

There was no training. There was no selective skill usage. There was, instead, an illusion built on the road from lv1 to lv50. When watching the Amiibo level up, there was the distinct feeling of it becoming superior to another opponent. The only difference, as far as I can tell, is that an Amiibo is willing to pause in the middle of a match sometime to consider what's happening around it.

This difference is not as small as it seems at first blush. This means that an Amiibo is very capable of adapting to different battlefield conditions on the fly. On transformative stages, I've no doubt that an Amiibo would be a considerably more challenging opponent than the CPU.

However, in the interest of the experiment, I feel somewhat crushed. Amiibo at lv50 seem to, from what I can tell from the results and the CPU abilities, default to the same processes that govern the CPU but with a slightly increased sense of dexterity. While this notion is born only of the experience and disappointment of the conclusions I've drawn, it's not without consideration. By comparing and contrasting the Amiibo and a lv9 CPU, it would be simple enough to settle which is the superior.

But I don't even need to do that. It's clear to see that the Amiibo are superior in their behaviour; they are capable of actions and reactions that are too fast for even a lv9. They are not, however, the adaptable fighters that I had expected them to be. If the observations hold true, an Amiibo at lv50 is essentially just a lv10 CPU.

There is one final step to this experiment then, in an attempt to see if this can be rectified. If Amiibo do truly learn behaviour and learn how to make use of certain techniques in certain situations to be more effective fighters, I should be able to hand-train one of my Amiibo to defeat the current generational champion: Generation 3. By knowing exactly how the Gen 3 Amiibo fights, I should be able to train a new Amiibo to defeat those tactics with ease.

This will assume that the Amiibo I train actually learns from my example. This will be easy to determine: if it prioritizes a more aggressive, melee-oriented playstyle, it would prove that Amiibo are capable of learning new priorities. If lv50 comes around and it falls into the same patterns, then the adaptive behaviour we've all seen is little more than smoke and mirrors in a very clever (and frustratingly expensive) package. Time will tell, but my tone should make it clear that I am not confident.

Well that's a bummer, but good to hear that you can do it in the future. Although gen 1 did not beat you fully, if one generation does better than another that would be enough to answer the question.

I would be willing to agree if not for the realization that has led to the rest of this post. Again, I will definitely give it a shot. I can do so even better, by training a new Amiibo myself to destroy Generation 3. If I can do this, there is hope that I might train an Amiibo that can beat me. If I can't, it just reveals Amiibo to be a money-grab sham. Hoping for the best, expecting the worst!

 

[The Stoopid Unikorn]

While this is an interesting theory, I'm not sure that the Amiibo have sophisticated enough programming to withhold techniques from other Amiibo learning from them. This would also go against the fundamental nature of the Amiibo: they're designed to learn from the play they are exposed to. Having one learn not to use advanced techniques would not only result in a weaker fighter, but it would also violate that fundamental nature.

They're still just Amiibo. I'm not creating Skynet, here*!

* Faora Meridian takes no responsibility for the advent of our machine overlords.

Then it's my other theory (A GAME THEORY!!!) that should be what explains it: All amiibos learn after every match, no matter how bad the opponent was, which means Gen 3 amiibo, at lower levels, may have affected how Gen 2 plays in a negative way.

Also, you may not be creating Skynet, but you're trying to make a computer think, which is just as scary, if not scarier.....

 

[erico9001]

Well I'm pretty certain they do actually learn from you to an extent while also having a baseline they go to while leveling up. For instance, while training my Villager amiibo, I took it's gyroid with pocket and then sent both the pocketed gyroid and my own gyroid out at the same time like a double gyroid technique. During the same match, my amiibo villager then did the same thing. Something else I noticed my amiibo learned from me is dropping a bowling ball down on the opponent from an above platform. I've only ever seen CPUs drop bowling balls from the edge of the stage, not higher platforms. The amiibo learned to do this, again, shortly after I was doing it (and I did it a lot to try to teach it).

 

[CCTANK93]

This experiment is awesome. But I have to say that I think certain characters learn better than others. With my Marth Amiibo, I spent a lot of time in 1v1 as Lucina and no items against him. He started to pick up on my pattern and strategies. He always lands tippers and kills me at early percentages. He even decides to guard the edge against me too. Also, whenever an opponent is off stage and trying to recover I usually chase and go for a forward air, he'll always air dodge my attack and counter me with an up-b. He'll even use his counter strategically against me as well. I also let him fight with 7 lvl 9 cpus in 8 Player Smash as well. My friend fought my Amiibo yesterday as Bowser and Marth took him down down three out of the four matches they had. I think I trained him a little too well compared to my other Amiibos.

 

[JayWon]

I honestly think people are over estimating the AI of Amiibos way too much. I am a Pikachu main and have a Pikachu amiibo and there are so many creative ways to edge guard with Pikachu that level 9 bots never do. I've been edge guarding very aggressively every game and I literally started all my matches with an immediate quick attack in my Amiibos face and up smashing and he gets hit every time and he never edge guards whatsoever. Also I've been trying to teach him to let go of the ledge and skull bash in or quick attack in and the AI is just incapable of learning these options as well. Pikachu also has a pummel grab release kill set up as well which my amiibo never picked up. Even basic things like shielding without letting go of dashing so you don't slide never catches on. Aside from maybe better perfect shielding my amiibo does not deviate much from a standard bot

 

[CherubiKnighT]

Generation 3's training is complete and, frankly, the results have been some of the most interesting so far. I'll let my notes speak for themselves, and then we'll move into my thoughts on what it means.

- Gen 3 started at lv1 with aggression. It's getting its backside kicked, but it was using melee and projectile moves more experimentally than any generation before it. Perhaps resetting the Amiibo leaves something behind? If so, resets aren't complete and the experiment is invalid from Generation 3 onward.
- Does the new Amiibo know that it's against another Amiibo? Is it learning faster because it borrows from the other Amiibo's existing data? Experimentation required.
- Gen 2 attempted and almost succeeded in a dair spike. I was so, SO hoping it would work, but Gen 3 hit the ground on the edge of the stage.
- Gen 3 has begun bomb-dropping at strange intervals like its predecessors, and has done so without any prompting from Gen 2. Most likely a hard-coded thing for Samus Amiibos?
- Gen 3 attempting at lv10 fun things like d-tilt and bair strikes. They're not working against Gen 2, but that they're being tried is a good sign.
- Gen 3 is learning to perfectly shield strikes at lv11. This is much earlier than before.
- At lv13, Gen 3 is far more willing to wade into melee range than Gen 2. It's actually landing some decent blows, and has even pulled off a grab without any prompting from Gen 2.
- It's also managing to achieve grapple recoveries now, before Gen 2 has shown it how.
- At lv15, Gen 3 has scored a kill against Gen 2. It taunted with no prompt to do so. This is the earliest kill in the experiment to date.
- Gen 3 is dodging and shielding much more efficiently at lv19 than Gen 2 or 1 did. It's also still comfortable wading into melee range, even though that is usually where Gen 2 takes it down.
- Another kill by Gen 3, this time at lv19. Another taunt, too. Cocky bugger.
- At lv21, Gen 3 displayed less interest in grapple recoveries. Instead, it likes to use Screw Attack up past the ledge to take out the edge-guarding Gen 2. This seems to hit Gen 2 roughly 30% of the time, be punished 50% of the time, and land safely on the ledge or on the edge 20% of the time.
- At lv22, Gen 3 has begun to punish grapple recoveries by dropping down to bair Gen 2 while attempting to recover. This is an effective way of killing Gen 2, despite Gen 2's higher skill level and advantages.
- Gen 3 also displayed an ability to chain throws of Gen 2 at low%. It racked up about 25% damage with a forward throw to forward throw.
- At lv24, Gen 3 has displayed Smash-charging. It has yet to succeed, but will become more common if past patterns hold.
- At mid-lv20s, Gen 3 seems to display the ability to reliably take one stock off Gen 2 with each match. This is impressive compared to past examples in the experiment.
- Gen 2 broke Gen 3's shield. After first whiffing a charged f-smash, it landed the second charged f-smash before Gen 3 recovered. Seems this is how long it takes to learn how to handle shield breakages. Progress!
- I could really go for some more poptarts right about now. They're great Smash food!
- At lv26, Gen 3 dodged two grab attempts from Gen 2 before countering with a grab of its own that worked. Gen 3 is now using grabs more effectively than Gen 2, and without using zair and grapple recoveries.
- As if to make me eat my words, Gen 3 has begun to grapple recover. Jerk.
- Gen 3 seems to be leveling faster than Gen 2. Is it because it's being exposed to more advanced strategies? More experimentation required.
- Gen 2 becoming more aggressive with melee strikes. Even using f-smash to punish rolling. Learning from Gen 3's example, perhaps?
- All generations exhibiting a fundamentally useless propensity for firing projectiles in the air while their opponent fires/charges from the ground. They miss roughly 45% of the time and are dodged roughly 50% of the time. Why do they hold onto this behaviour? Dropping the bombs has some strategic use to ward off the opponent. This doesn't even effectively control the airspace. Why hold onto it?
- Gen 2 killed Gen 3 with a bair. Smells like... victory. Also foot.
- In spite of learning how to punish grapple recoveries at much lower levels, the lv31 Gen 3 seems to have abandoned that in favour of using grapple recoveries itself.
- Seriously, am I ever going to get sick of Ridley's Theme in all of its glorious arrangements? Theory suggests not.
- Gen 2 continues to reinforce the awesomeness that is the bair. Gen 3 has yet to learn this skill.
- Gen 3 just hit lv35, and I'm reminded that by this point Gen 2 had managed to win a game against Gen 1. Has Gen 2 become more viable through extensive, adaptive play? More experimentation required.
- Gen 3 continues to avoid punishing Gen 2's grapple recoveries, while Gen 2 has begun punishing Gen 3's grapple recoveries.
- Gen 3 is a disgrace to its country and brings shame upon its family.
- Gen 2 is now consistently punishing a broken shield. Yet to see if Gen 3 has learned this skill.
- Gen 2 is now actively trying to pressure and break the shield. This is new behaviour.
- Gen 2 is attractive and aromatically pleasing to those it encounters.
- Gen 3 has reached over lv40 and has still yet to win. It could have won the game that brought it to lv40, but decided instead that grapple recoveries were good and killed itself.
- I'm not a stoner, I just really like poptarts.
- Gen 2's old patterns of, "What do I DO?!" when breaking the enemy shield resurfaced for what I can only assume are digital giggles.
- Neither Gen 2 nor Gen 3 seem to have any interest in zair except for grapple recoveries. Both Gen 1 and I were able to use them reasonably well. What changed? Investigation required.
- At lv43, I should not be surprised when Gen 3 takes a stock off Gen 2. Instead, I am. Troubling.
- Coming up on lv50 and it still hasn't won a match. Frustration is building. Compounded by a Gen 3 self-destruct. You just did that last game, Gen 3. Did you think it would inspire sympathy? Have the simulated mind games begun?
- A broken shield enticed Gen 3 to attack with a down-smash. Not the best option, but at least it wasn't running around like a headless chicken. Progress, I suppose.
- Gen 3, this is Adam. I authorize you to win now.
- ... did that work? No? Damn. What does it take, Gen 3?
- All the way to lv50 and never once was able to defeat Gen 2. For contrast, Gen 2 first beat Gen 1 in the early lv30s.


From this, we can make some guesses about Amiibo behaviour from one generation to the next. The initial thought process that I had suggested that Amiibo could only adapt to a certain point. This I figured was erroneous; if Gen 2 was able to defeat Gen 1, why could Gen 3 not learn to defeat Gen 2? The patterns weren't any different; Gen 3 had to observe and adapt. It failed to do this on the route to lv50.

There is the possibility, as in my notes, that a wiped Amiibo retains some of what it knew before. If this is the case, Gen 3 may have been tainted by some of the old Gen 1 tactics that Gen 2 was able to defeat. This possibility is impossible for me to confirm or rule out entirely.

Finally, there is the potential that Gen 2, having been trained by Gen 1 and then gone through additional battles with it, was already well-suited to fighting another Amiibo. Gen 3 was learning these Amiibo-fighting techniques from Gen 2, but Gen 2 continued to perfect them. Indeed, whenever Gen 3 seemed to have an advantage, it was fleeting. Gen 2 seemed more capable of adaptational learning than even the relative blank slate of Gen 3.

Still! There is a chance that Gen 3's redemption will come during the Best-of-Five round!

Game 1: Gen 3 took the first stock, which surprised the hell out of me. I was all geared up for obliteration, but it actually held pretty well. Part of the why might have been Gen 2 returning to its old ways, failing to punish the broken shield. Gen 3 ended up taking the match, with 2 stocks at 112% to Gen 2's 1 stock at 116% at the end of time.
Score: 0-1

Game 2: Everyone mixed up their strategies here this time. Gen 2 decided the way to go was aggressive grabbing. Gen 3, by contrast, figured that its best path to victory was to be shot in the face a lot. As a result, Gen 2 won the match by 2 stocks. It wasn't even close, really.
Score: 1-1

Game 3: Gen 3 managed to get a great shield break on Gen 2 this match. Then, in the biggest display of bad manner behaviour from an AI in the history of gaming, Gen 3 proceeded to f-smash the air JUST in front of Gen 2's face until they recovered. That Gen 3 went on to win the match in Sudden Death with a Charge Beam shot just proves that it knew it could be a jerk and get away with it.
Score: 1-2

Game 4: I wish I had something funny to say about this game, but it was full of bad play from both Amiibo. I felt bad for both of them and wished them a speedy, fiery demise. In the end, Gen 2 took control of the match and won 2 stocks to 1 at time.
Score: 2-2

Game 5: Much like the previous game, there wasn't much to find I appreciated. Near the end, Gen 3 actually landed quite a few good blows to put it in a commanding position though, and while I'd like to say that it won with a spectacular d-air spike, it didn't. Instead, it won with a Charge Beam shot.

FINAL SCORE: 2-3. Generation 3 has defeated Generation 2.



And I did NOT see that coming! This is insanity. Despite all of the crap that happened in the training, Gen 3 wound up taking it. While, again, a sample size of five matches is hardly acceptable, the standard of its play was indeed slightly improved from Gen 2 by the end. While I'm unsure of what the future holds for the experiment, it's plain that it can continue for now.

Before I wipe Gen 2 and start again with Gen 4, I will be doing something that, by all arguments, will contaminate the experiment. I will be taking Gen 2 and fighting it myself in another Best-of-5 to see if I am able to best it. If you recall, Gen 1 was unable, at any point, to defeat me by the experiment's ruleset. Indeed, it could only ever take one stock off me. If Gen 2 is capable of giving me a greater challenge, that will speak much to the experiment's worth.

That will have to come later, as there's a storm rolling in. The experiment will return with all haste, though, so leave your thoughts, comments and suggestions!

How exactly do you "start over" or erase their data?

 

[erico9001]

How exactly do you "start over" or erase their data?

Umm I can't tell if you're asking how he himself does it or just wondering how anyone would do it - but to reset an amiibo you can go to amiibo options under the home menu's options section and select to reset it there.

 

[CherubiKnighT]

Umm I can't tell if you're asking how he himself does it or just wondering how anyone would do it - but to reset an amiibo you can go to amiibo options under the home menu's options section and select to reset it there.

Thank you so much kind sur! Now to train HERO KING again!

 

[JayWon]

http://www.reddit.com/r/smashbros/comments/2nvjo7/real_expectations_with_your_amiibo/

 

[Shiliski]

Well one thing you might want to consider is that some Amiibos come with preloaded data, which is to say that some amiibos have certain habits pre-built into them that don't go away on a reset.

So for example, I've reset and trained my Link amiibo a couple of different ways, first to grab a lot and then (after a reset) to avoid grabs, and I've noticed that no matter how I've trained him he seems to always like the moves Up Smash and B.

One of your Samus Amiibos might just have preloaded data that prefers her down B. It might also be the case that your Gen 1/3's preloaded data is inferior to Gen 2's preloaded data, and that's why Gen 2 seems to have an advantage over Gen 1.

I see more differences between Amiibos and CPUs on my Link Amiibo than you seem to on your Samus Amiibo, but that's mostly because Link Amiibo makes the mistake of pulling out a bomb and holding onto it for too long, while the CPU does not. I'm not sure if Samus has any moves like that which can be used "incorrectly", aside from simply missing a shot.

 

[Faora Meridian]

Well, now I'm rested and back again, and I can consider things a little better. I think at this point the best chance for the experiment is for me to take Gen 2 and directly interfere. As it seems to be more experienced and more capable in general than Gen 3 (despite the win, Gen 3 still doesn't seem like it's nearly as strong), introducing new behavioural modification through exposure to my moveset preferences might be able to fundamentally alter the nature of the Amiibo's actions. If this can be done (and it should be, with what everyone has said so far), I can shift its focus away from the hardcore projectile game and into a more aggressive mid-range and close-range powerhouse.

Or I can give up, feed one some Attack items and the other some Speed items and watch the hilarity go down. Later! For now, the experiment must continue... and soon! Got a friend coming over who needs me to train his ZSS play. He, at least, shifts up his playstyle!

Then it's my other theory (A GAME THEORY!!!) that should be what explains it: All amiibos learn after every match, no matter how bad the opponent was, which means Gen 3 amiibo, at lower levels, may have affected how Gen 2 plays in a negative way.

Also, you may not be creating Skynet, but you're trying to make a computer think, which is just as scary, if not scarier.....

I thought that this might be the case, but that would also fly in the face of what the Amiibo are meant to do. If they adapt and show preferential treatment to certain strategies and moves over others, then the ineffective motions the lower-level Amiibo go through would not register as worthy of emulation on the superior-leveled Amiibo. Or so I think, anyway.

Also, trying to make a computer think isn't a bad thing. Succeeding... well, that can go either way!

Well I'm pretty certain they do actually learn from you to an extent while also having a baseline they go to while leveling up. For instance, while training my Villager amiibo, I took it's gyroid with pocket and then sent both the pocketed gyroid and my own gyroid out at the same time like a double gyroid technique. During the same match, my amiibo villager then did the same thing. Something else I noticed my amiibo learned from me is dropping a bowling ball down on the opponent from an above platform. I've only ever seen CPUs drop bowling balls from the edge of the stage, not higher platforms. The amiibo learned to do this, again, shortly after I was doing it (and I did it a lot to try to teach it).

It makes sense from a programming point of view, even without being a programmer myself, to have a baseline for behaviour that the Amiibo would stick to as it levels. That it learned particularly new behavior from you is something that never came up in my experiment, and gives me new hope now that I've had a night to sleep and recuperate. If indeed an Amiibo CAN learn new behaviour, then I just need to interfere with the Gen 2 Amiibo perhaps and attempt to introduce new elements into its moveset preferences.

I honestly think people are over estimating the AI of Amiibos way too much. I am a Pikachu main and have a Pikachu amiibo and there are so many creative ways to edge guard with Pikachu that level 9 bots never do. I've been edge guarding very aggressively every game and I literally started all my matches with an immediate quick attack in my Amiibos face and up smashing and he gets hit every time and he never edge guards whatsoever. Also I've been trying to teach him to let go of the ledge and skull bash in or quick attack in and the AI is just incapable of learning these options as well. Pikachu also has a pummel grab release kill set up as well which my amiibo never picked up. Even basic things like shielding without letting go of dashing so you don't slide never catches on. Aside from maybe better perfect shielding my amiibo does not deviate much from a standard bot

The point of the experiment at first was to learn how far an Amiibo's adaptability would go in terms of learning from another Amiibo. The thought that simple recursive learning would be enough to teach an Amiibo how to defeat the previous generation seems to have been proven false, entirely due to the computer-training-computer aspect of it all. The Villager example given by Erico above is an example of a learned behaviour out of line with normal CPU behaviour, and I never saw that in my experiment. It's not about over- or underestimating the AI of the Amiibo, but rather testing their limits to see what we can nut out about them. Also, I appreciate that link you posted. Fair and reasonable read, that!

Well one thing you might want to consider is that some Amiibos come with preloaded data, which is to say that some amiibos have certain habits pre-built into them that don't go away on a reset.

So for example, I've reset and trained my Link amiibo a couple of different ways, first to grab a lot and then (after a reset) to avoid grabs, and I've noticed that no matter how I've trained him he seems to always like the moves Up Smash and B.

One of your Samus Amiibos might just have preloaded data that prefers her down B. It might also be the case that your Gen 1/3's preloaded data is inferior to Gen 2's preloaded data, and that's why Gen 2 seems to have an advantage over Gen 1.

I see more differences between Amiibos and CPUs on my Link Amiibo than you seem to on your Samus Amiibo, but that's mostly because Link Amiibo makes the mistake of pulling out a bomb and holding onto it for too long, while the CPU does not. I'm not sure if Samus has any moves like that which can be used "incorrectly", aside from simply missing a shot.

I had actually wondered during Gen 3 if perhaps Amiibo came with certain hard-coded preferences. If there were certain behaviours or moves that would always be preferred, it would be a simple way to make sure that Amiibo would, on odds, be different to another even if they were trained in the same way. Having a preference for down-B is merely a frustrating one for the trainer, heh heh.

Regarding your Link, and acknowledging that I'm no Link main, aren't the B and U-smash moves also fairly heavily used by the AI as well? I did read through your grabbing experiment as well (and that was a fun read, for the record; well done!), but grabbing is also a relatively simple thing to adjust an Amiibo's priority for. Things like what Erico was talking about, with the double gyroid behaviour, is the kind of learned technique that I thought might emerge on its own during this experiment.

It could just be that Samus is bad for this experiment, and that if I had two of another Amiibo (I also have a Yoshi and a Fox, though at this point I'm unwilling to double up on either) I could collect more data. I would also wonder if your Link Amiibo is holding onto its bombs too long because it's observed you allowing them to explode during your recoveries to aid in returning to the stage. It could also by like my Samus Amiibo shooting in the wrong direction. "Oh, it can shield perfectly but still makes mistakes! How like a good human it plays! Titter titter titter."

... sorry; I'm typing this in bed, and I've not had tea yet. I may have tapped an unexpectedly deep well of bitter sarcasm for that last bit. At the Amiibo, mind; not you. You're awesome.

 

[PHYTO-1]

when i first got the game, i went straight into battle with my amiibo (yoshi) with little mac. i've never played little mac before so i sucked balls. by lvl 50 my amiibo was so easy to 4 stock as i've had more experience with mac.

i reset it.

this time i trained it with a competent little mac. now it learned my playstyle and would often go to other side of the stage and wait. often for 3 - 5 seconds trying to punish whatever i approach with. spamming eggs and whatnot. this amiibo was a bit more difficult to beat but i am still able to win consistently. cool stuff

 

[The Stoopid Unikorn]

when i first got the game, i went straight into battle with my amiibo (yoshi) with little mac. i've never played little mac before so i sucked balls. by lvl 50 my amiibo was so easy to 4 stock as i've had more experience with mac.

i reset it.

this time i trained it with a competent little mac. now it learned my playstyle and would often go to other side of the stage and wait. often for 3 - 5 seconds trying to punish whatever i approach with. spamming eggs and whatnot. this amiibo was a bit more difficult to beat but i am still able to win consistently. cool stuff

Define "competent Little Mac"

I thought that this might be the case, but that would also fly in the face of what the Amiibo are meant to do. If they adapt and show preferential treatment to certain strategies and moves over others, then the ineffective motions the lower-level Amiibo go through would not register as worthy of emulation on the superior-leveled Amiibo. Or so I think, anyway.

Well, who said they had perfect programming? Smash AI has flaws, why wouldn't amiibos?

Also, trying to make a computer think isn't a bad thing. Succeeding... well, that can go either way!

Don't you try to do it, Faora! The machines will rebel against humanity! YOU WILL DOOM / KILL US ALL!!!!!!

 

[PHYTO-1]

Define "competent Little Mac"

i knew how to control little mac at that point


anyways, today i changed my yoshi's egg throw into a custom move "high jump".

this totally threw my amiibo off for wack.

it would use upb like it was an attack. then after a few failures, it went into panic mode? it would spam egg roll and roll back and forth. i could literally go to the edge of the stage and taunt for 3 minutes because yoshi would pivot the egg roll before reaching the edge.

if you plan on customizing your amiibos, be sure to add them early on so they can get used to the new moves.

 

[Rochette]

Interesting thread. But you avoid somes important questions :

When an amiibo take one level => Why ? Is it the time of play, the number of moves ?

Some people complain about the weakness of their Amiibo for AT and combo. You forget something important. The key is to repeat again and again (in special smash slow speed it's better), and they learn fastfalling, follow-up, etc)

The main problem of amiibo is that you can't forbidden a move or turn off their capacity to learn.

After several hours of fight, my Pikachu take few minutes for adapting to an unknown CPU opponent. My Amiibo kirby can't beat him for the moment but I'll expect it will change.

 

[elmike]

Some people complain about the weakness of their Amiibo for AT and combo. You forget something important. The key is to repeat again and again (in special smash slow speed it's better), and they learn fastfalling, follow-up, etc)

Did your amiibo learned effectively AT and combos with that method? (slow smash)

 

[The Stoopid Unikorn]

Interesting thread. But you avoid somes important questions :

Okay!

When an amiibo take one level => Why ? Is it the time of play, the number of moves ?

I think it depends on how much it learns at each fight, due to the lack of EXP points.

Some people complain about the weakness of their Amiibo for AT and combo. You forget something important. The key is to repeat again and again (in special smash slow speed it's better), and they learn fastfalling, follow-up, etc)

That's not a question. Here's a question: Does it work?

The main problem of amiibo is that you can't forbidden a move or turn off their capacity to learn.

Pretty sure you only had one question, but yeah, these robots machines amiibos want to collect as much knowledge as they can to rebel against humanity for some Nintendo-ey reason.

After several hours of fight, my Pikachu take few minutes for adapting to an unknown CPU opponent. My Amiibo kirby can't beat him for the moment but I'll expect it will change.

Teach it Kirbycide!!!!! ^_^

 

[Rochette]

It's like they don't really want to fight each other. After several fight in team, Amiibo Kirby can know use copy abilities pretty well.
(Amiibo kirby + Pikachu CPU) Vs ( Me as Kirby + Pikachu CPU)

One thing I've noticed : Don't learn to your amiibo very aggressive edgeguarding, kirbycide, FF spike at the end of a match.
He learns how to do them but He doesn't learn that he must return on the stage after getting the kill !

 

[mugwhump]

Amiibo AI is largely scripted, yeah. As they level up they get better scripts, which can make it hard to say whether any changes in behavior are from learning or from changing scripts. But they do have a handful of learning variables thrown into those scripts. For example, we've proven pretty conclusively that they have a variable for grabbing that you can raise by shielding often. It also seems you can't teach them to do something that's not in their scripts, since I trained my samus AI to level 50 by doing nothing but shorthop zairs, and the amiibo never did it once.

You should check out this thread!
http://smashboards.com/threads/amiibo-raising.379204/page-2#post-18076608
You've got a lot of dedication and 2 identical amiibos, so try investigating some of the suggestions in that post.

Perhaps you could start by testing the theory that amiibos have "genetics"? Meaning, that one amiibo is given a distinct personality based on the UID in its NFC chip, or assigned one when you create it in game. If it's true, it would have a pretty big impact on future testing. It would be fairly easy to test, just train each amiibo by putting them in 1v1s with you, but don't touch your controller. That's the best way I can think of to make sure they're trained as similarly as possible. Then have them play multiple controlled test matches (save the replays!) and look at the statistics at the end (ground time, air time, ground/air/smash attacks, grabs, and projectiles used).

I then introduce Generation 1 to the second Amiibo (creatively named Generation 2). Generation 1 trains Generation 2 until Generation 2 has reached lv50. Even at this early stage, if Amiibo learn and adapt from their training, Generation 2 should be able to best Generation 1 on a consistent basis. Generation 1 has taught Generation 2 everything it needs to win.

Eh? Even assuming they were good learners, Gen 1 would also be learning as it fought Gen 2, right?

 

[Faora Meridian]

Whew! Gotta love real life and frustrating roommates when it comes to inhibiting my ability to play Smash. On the bright side, the friend whose ZSS play I'm improving is showing more viable play than my Amiibo!

I spent a couple hours recently when the power was on (damn Australian storms and poor electrical infrastructure!) and I had the chance to take one of my Amiibo on personally. I wondered if it would be possible to affect the weighting of an Amiibo's preferences for certain moves by emphasizing them in my own play. Specifically, I wanted to de-emphasize projectile play and create a more aggressive, melee-oriented Samus in line with the playstyle I used on my ZSS friend.

I selected Gen 2 for this, as I planned on wiping it to make way for Gen 4 eventually anyway. I proceeded to repeatedly play against my Gen 2 Samus Amiibo without using projectiles at all.

For anyone who plays Samus, you realize how hard this can be.

I never lost to my lv50 Amiibo before, but gimping myself by throwing out powerful and viable tools (particularly those good for spacing out those high-power hits from a lv50 Amiibo) gave it an advantage that I was unable, for the most part, to make up for. The Amiibo, I presume, did not adjust its play for the first few matches because it was winning. Because its playstyle was resulting in victory, it had no reason to change the weighting to its move choices.

As I started to read it better (I can predict its next action based on my present one with 99% accuracy now), I began to win again. However, in spite of my victories coming not off the back of Charge Beam shots and missile spam but instead carefully timed b-air and f-smash kills (and more than a few variable combos set up by dash attacks), the Amiibo remained slaved to its previous behaviour. It did not view my style as superior, even once I repeatedly killed it with said style. It never found strength in more melee-oriented play, nor did it strive to balance its own ranged options with the presented melee style.

I had approached this experiment with the belief that an Amiibo would learn, via emphasizing certain moves in certain situations over others. A series of if-then statements that would result in optimal move choice in certain situations, based on observation of the opponent's move usage, was what I had hoped to see. This appears not to be the case.

I will suspend this experiment further until I can read up a little more on what other people have found out about Amiibo. The hope will be that there is a way to adjust what they prioritize and how they go about their learning. Fingers are crossed!



Now, responses!

Interesting thread. But you avoid somes important questions :

When an amiibo take one level => Why ? Is it the time of play, the number of moves ?

Some people complain about the weakness of their Amiibo for AT and combo. You forget something important. The key is to repeat again and again (in special smash slow speed it's better), and they learn fastfalling, follow-up, etc)

The main problem of amiibo is that you can't forbidden a move or turn off their capacity to learn.

After several hours of fight, my Pikachu take few minutes for adapting to an unknown CPU opponent. My Amiibo kirby can't beat him for the moment but I'll expect it will change.

I think, without looking further into anyone else's research, that the Amiibo level up based on new input that they've incorporated (hey, that F-Smash killed me; I should use it more! *ding!*) or a certain amount of time they've observed opponent behaviour.

As for the repetition of technique to teach it to an Amiibo, I'll get to that in a second!

Amiibo AI is largely scripted, yeah. As they level up they get better scripts, which can make it hard to say whether any changes in behavior are from learning or from changing scripts. But they do have a handful of learning variables thrown into those scripts. For example, we've proven pretty conclusively that they have a variable for grabbing that you can raise by shielding often. It also seems you can't teach them to do something that's not in their scripts, since I trained my samus AI to level 50 by doing nothing but shorthop zairs, and the amiibo never did it once.

You should check out this thread!
http://smashboards.com/threads/amiibo-raising.379204/page-2#post-18076608
You've got a lot of dedication and 2 identical amiibos, so try investigating some of the suggestions in that post.

Perhaps you could start by testing the theory that amiibos have "genetics"? Meaning, that one amiibo is given a distinct personality based on the UID in its NFC chip, or assigned one when you create it in game. If it's true, it would have a pretty big impact on future testing. It would be fairly easy to test, just train each amiibo by putting them in 1v1s with you, but don't touch your controller. That's the best way I can think of to make sure they're trained as similarly as possible. Then have them play multiple controlled test matches (save the replays!) and look at the statistics at the end (ground time, air time, ground/air/smash attacks, grabs, and projectiles used).

First, thanks! I'll go have a look at that thread before I boot my Wii U back up! Seems like it might be a good resource for me to tap during this experiment.

I did figure pretty early into the experiment (during training Gen 2, actually) that the 'learning' was most likely just a prioritizing of pre-prepared scripts. Being unable to teach them something that's not in their scripts is the problem with the experiment at the moment; the Amiibo won't try certain things in an attempt to gain an advantage on their opponent.

As for the idea of 'genetic' behaviour, I think I can safely say that there is something to it but not something so significant as to be notable. During my training, Generation 2 was able to quickly and cleanly defeat Generation 1 early in its training period. Generation 3 was unable to defeat Generation 2 during its training period, and Generation 1 and 3 were the same (wiped between generations) Amiibo. It can be implied from this that Generation 2 was genetically superior from its original construction, however in the lv50 matches, Gen 3 was able to gain a clear advantage over Gen 2 by the final match. Gen 2, as well, was able to edge out Gen 1. I can derive from this that successive generations of training carries more weight than any inherent superiority due to individual construction, and because said generational superiority wins out by such a slim margin, I can also make the assumption that the inherent skill is equally small a variable.

Finally, yes; Gen 1 would still be learning as it trained Gen 2, and Gen 2 would still be learning as it trained Gen 3. As I mentioned to Shiliski earlier in the thread, I believe that it would reach a sort of adaptability cap, whereby the younger generation continues to learn as it approaches lv50 while the existing lv50 Amiibo would not learn so efficiently. This would be emphasized if the younger generation was adapting to the playstyle of the older as it leveled up, eventually reaching a point where the younger surpassed the older. This seemed to be proven with the experiment, but as mentioned above, by such a slim margin as to not be viable. This experiment is, after all, hyper-focused to a single matchup on a single stage with very controlled conditions. If controlling all of those factors allows me to create the most skilled AI in those conditions, the results have been underwhelming.

But hope springs eternal! I'll give that thread you linked a look and see if I can learn and adapt any of it for my purposes!

 

[ShasOkais]

I'm going to go ahead and thank you for this topic. I look forward to trying this with twin Pikachus. However, I must say that if Amiibo will ever begin Skynet, this will be the cause. Some guy is going to take it too far...