Amiibo Theory - Generational Learning

[The Stoopid Unikorn]

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...

Would that make Fils-A-Mech a Terminator?

 

[ShasOkais]

I thought I might give this a shot, but after acquiring a second Pikachu Amiibo, my results have been fairly discouraging. Gen 1 has consistently beaten gen 2 in every single game. They both are lvl 50, no stats, no effects, and no custom specials. Every match has been 30 stock on Final Destination, no items. Gen 1 wins with an average of 7 stock remaining each time. I'm honestly having a lot of trouble making sense of this, and I'm really beginning to question if/how Amiibos actually "learn". I'm also afraid to reset either Pikachu until I figure out whats going on here.

If anyone knows what is going on with my Pikachus, I'd be happy to know, and any sensible answer would be greatly appreciated.

 

[Funkermonster]

I don't have enough time to read everything from all those lengthy, yet informative posts but I have a Question: Is it me or do most Amibos rarely seem to ever try to jump offstage and gimp people? And they also seem to use Smashes very often even at low percents.

Before I reset it , my Pikachu amiibo just sat there onstage waiting for me to come back so he could kill me with a smash attack, and he never attempted to kill me offstage with aerials (particularly the Dair) even once. I reset him to try and fi some bad habits like that, and I haven't yet gotten him to level 50.

I've seen videos of other amiibo (Peach, DK, and Kirby) who also never attack offstage.

I haven't gotten my new Yoshi to level 50, but I've heard complaints from others about him just camping with eggs most of the time and smashing anyone who comes near. Unlike my Pikachu, I have at least one memory of him killing a CPU offstage the last time I played with it.

Or maybe they just take a longer time to learn stuff than I think they do?

 

[yogo540]

Over time, your Amiibo will fight with less human tendencies. The Amiibo algorithm doesn't create better fighters, it creates similar fighters.

So, you're making a copy of a copy. The problem is that each generation replaces some of that "human" data with "Amiibo" data. They don't get better, they just become more generic.

Each new generation will only become better if, "by luck", it picks up a higher proportion of good habits to bad habits.

 

[Faora Meridian]

So, just a heads up for those interested in the continuation of this experiment, good news and bad news! The good news is that I will be working on it at last again tomorrow, by resetting Gen 2 completely and training it up myself to use techniques that are not apparently hardwired into the Amiibo learning system. The bad news is that I'll be doing this because the clutch on my car has decided that now is the perfect time to fail, and as a result of that I probably am also unemployed. Pros and cons are all relative!

I'm also very interested in the results of the above Pikachu experimentation should it move to another generation. I will be interested to see if Gen 1/Gen 3 maintains its edge, or if in training it Gen 2 learns how to defeat it. We shall see.

And if in the end it is determined that the experiment has failed, then I shall stick equipment on them and pit them in fights to the death for my amusement and profit! And then, nothing shall stop me!

Anyway, stay tuned!

 

[Trekkerjoe]

Your car broke down? I hope it is not as horrible as you make it out to be. I hope you can get your car fixed before it jeopardizes your job. I also hope to hear what you come up with in this experiment, I just hope it is under better circumstances.

 

[Faora Meridian]

Your car broke down? I hope it is not as horrible as you make it out to be. I hope you can get your car fixed before it jeopardizes your job. I also hope to hear what you come up with in this experiment, I just hope it is under better circumstances.

Hey! Actually, it was worse; the clutch was gone, which cost me a buttload of money to replace, and while there they discovered that there was a crack in my radiator that was going to cause completely failure in short order... and that added to my bill, of course. They also found some non-threatening problems that I, for financial reasons, am going to let slide for the moment. But it all came out okay; the car's running fine now!

As for the experiment, it has been an... interesting road. This will likely be the last post to this thread for the moment until and unless I get new Amiibo pairs (which is unlikely). I think the results of the experiment allow certain assumptions to be made, and from there I think my work is done. Anyway, on to the experiment proper!

As was previously noted, the intention was now to take my old Gen 2 Amiibo, wipe it clean, and train it from the ground up to not make use of the problematic behavior inherent to the other Generations of Amiibo. Specifically, the goal was to try and abolish its reliance on projectiles in favour of Samus' powerful (if occasionally clunky) melee and aerial game. Here's the notes I took while training this New-Gen Amiibo!

- Ran me down with dash attacks at lv1. Aggressively attempted d-air and up-b attacks at lv3. Probably retained something from previous wipes.
- Used d-tilt without prompting at lv5. Began grapple-recovery without taunting. Off-ledge game consistant with normal ability for Amiibo at lv6. Still no ranged attacks by lv7; previous generations began using projectiles much earlier.
- At lv7, Amiibo began to charge Smash attacks; if memory serves, this is in advance of the previous generations.
- Used first projectile (a completely uncharged Charge Beam shot) at lv9, much later than previous generations.
- Accidentally tapped off a Charge Beam shot. Oops. Amiibo used the same tap-shot a couple more times around lv11. Still far less use than previous generations.
- At lv14, Charge Beam tap-shots are still very rare. Increased emphasis on air play. Amiibo seems to like to rely on up-B to deal damage aside from lucky Smash attack hits.
- As of lv15, the Amiibo learned on its own to charge the Charge Beam and fire missiles. Only one missile was launched the entire match, but I suspect this trend will continue.
- At lv16, Amiibo has commenced highly-effective air- and spot-dodging. Can't remember, but I think this is much earlier than normal, too.
- Embarassing. At lv17, the Amiibo hit me with a meteor smash right on the edge of recovery and took a stock. It then saw me lose another stock to my own stupidity. The Amiibo was still defeated, but I'm disappointed in myself.
- At lv17, the Amiibo has also begun to make more consistent use of projectile attacks. Beginning to think that the Samus Amiibo has no way to effectively alter the weight on it's projectile game compared to other avenues of attack. More investigation required.
- I hate not being able to use projectiles at all. This is frustrating.
- As of lv20, the Amiibo has begun to make use of Super Missiles.
- At lv21, the Amiibo took a stock off me through conventional means. This not-firing-projectiles thing is getting to me.
- lv23, another death on my part. Amiibo seems to be very good at meteor smashes this time around. I only had limited success against it with them in this cycle. Another hard-coded thing? Projectile use is steadily increasing, but still at lower levels than norm.
- Amiibo took two stocks off me in one game at lv25 and 26. On the one hand, I'm happy to be learning how to play when my projectile game is limited. On the other, I would REALLY prefer to use my projectiles.
- You know who else didn't have Charge Beam usage? Hitler. Look where that got him.
- Abusing predictable Amiibo behaviour off the edge isn't a good way to teach it. It IS a good way to get some frustration worked out.
- At lv28, old reflexes finally got the better of me and I fired missiles. Amiibo doesn't seem to have picked up the projectile usage from that; projectile use is actually down. Maybe not so hardwired as I thought?
- At lv30, Amiibo showed new learning. It has begun to try and lock me into up-air-to-up-air combos. This has not been used by previous Amiibo generations. Alternate script weighting?
- At lv32, the Amiibo seems to show a massively disproportionate proclivity for the up Smash. Potential counter to my air approaches. Requires further experimentation.
- As of lv34, the Amiibo shows only the slightest inclination toward projectile attacks. Maybe my suffering is not in vain.
- At lv35, the Amiibo continues to show a high degree of skill in meteor smashes, while I continue to show a remarkable level of sloth toward avoiding them.
- Air approach against the Amiibo is ill-advised. Dat up-Smash, man. Dat up-Smash.
- At lv36, the Amiibo has become a serious threat with my present playstyle. The growing bonuses to its damage and resistance to my blows are becoming problematic to overcome while limiting my moveset. Frustration grows, but the experiment seems promising at this point.
- At lv 37, the Amiibo has defeated me. It wasn't even close; it won by two stocks. The experiment shows more promise.
- I can still defeat the Amiibo even up to lv40. My ego feels better.
- Progression continues as anticipated right up to lv45.
- At lv48, the Amiibo defeated me again. The damage increases are just too much for me. It'll be interesting to see how it fares against Gen 3.
- Defeated one final time at lv50. Projectiles are still a rarity. Focus is on ground play with spikes. Impressive shift in focus from the Gen 1-3 examples.


I was surprised and impressed. Everything that I was able to observe from the Amiibo while training it was almost in direct opposition to everything that I had seen before. By completely de-emphasizing projectile play while training the Amiibo, I created an environment of extreme player frustration that reinforced the use of its other skills. Sticking to these same rules was excruciatingly painful, but it seemed to pay off in the end with an Amiibo at lv50 that was capable of crushing me regularly in these altered rules of engagement.

But of course, the real test comes from the Best-of-Five series. In this, the New-Gen Amiibo was pitted against the old victor: Generation 3. Gen 3 hadn't been used in any fights since it won its Best-of-Five against Gen 2, and I was interested in two things in this matchup. Firstly and most importantly, I was wondering how it would handle the more aggressive, melee-oriented playstyle of the New-Gen Amiibo. Secondly, I was wondering how long it would take the New-Gen Amiibo to start using projectiles as a result of observed behaviour and how extensive that use would be. I found my answers; check out my match notes below!


Game 1: Gen 3 took frst stock very early. Its projectiles are overwhelming the New-Gen Amiibo. The New-Gen Amiibo did begin using a few more projectiles as a result of this match, particularly after scoring a kill with a Charge Beam shot. Its fondness for up-smash often was overwhelmed by Gen 3's f-smash, but Gen 3's aerial approaches were almost always punished by it. Blast from the past: Gen 3 failed to punish a broken shield. At least the New-Gen Amiibo training seems to be more practical with its projectiles. Surprisingly, with the pros seemingly outweighed by the cons, the New-Gen Amiibo wins the first game by 1 stock (60%).
Score: 1-0

Game 2: New-Gen Amiibo's took the first stock very early. Its projectile usage starting to melt into Gen 3's usage, but somehow more effectively. Watched a whole minute go by with neither Amiibo being able to hit the other... very odd. New-Gen Amiibo WINS the second game by one stock (3%).
Score: 2-0

Game 3: Gen 3 took the first stock very early because of a poor recovery by the New-Gen Amiibo. It's picking up more of Gen 3's less-tasteful habits now and, with only a few exceptions, the two seem to be playing almost identical. The New-Gen Amiibo seems to maintain the advantage, though; even that early stock loss didn't slow it down much and it retook a slender lead with its second stock. Gen 3 reversed it in the end to WIN the third game by one stock (133%).
Score: 2-1

Game 4: New-Gen takes a hard-won first stock in the game three whole minutes in. Had to pleasure of watching the two Samus' f-air each other at the same time and recoil from the hit, only to do it again, also simultaneously, as they fell. My amusement was intense. Less amusing was the growing reliance of the New-Gen Amiibo on projectiles. On the other side, Gen 3 began to effectively incorporate some of the New-Gen's successful strategies (like air approaches covered with up-smash) into its own moveset. Was intrigued to see Gen 3 short-hop to bair three times in a row, even though New-Gen dodged back each time. In the end, New-Gen WON the game by two stocks (187%). If this were a real best-of-five tournament, New-Gen would have won at this point 3-1; a first in this experiment.
Score: 3-1

Game 5: New-Gen Amiibo took a very early stock off Gen 3 through off-edge play, much to my delight. Was surprised to see Gen 3 engage in some up-smash juggling; I would have thought that the New-Gen Amiibo (or any Amiibo at lv50) would avoid that situation. In spite of the New-Gen's growing reliance on projectile play, it's becoming clear that it's by far the more capable melee and aerial combatant. This seems to give it an edge over Gen 3 on a relatively consistent basis. At the end, New-Gen WON the game by one stock (3%).

FINAL SCORE: 4-1. The New-Gen Amiibo has defeated Gen 3.


There are a number of reasons why this might have been the overall result, some of which are extremely flattering regarding my superior, personal Amiiibo-training techniques (the secret is to hate everything about yourself while you deny yourself character tools that make for better gameplay). For those who would leap to that conclusion, I appreciate your faith in me!

However, I believe the answer is far simpler and inherent to the experiment itself. It has been shown in other threads that Amiibo that are trained by human opponents and Amiibo trained by lv9 CPU opponents are not equal and the human training wins out. I would theorize that, while the intention of the experiment was to create through generational learning a situationally-perfect Samus Amiibo, it was this very goal and method that produced an INFERIOR Amiibo.

Only half of the Amiibo generations in this experiment (Gen 1 and New-Gen) were trained by my hand directly. By contrast, Gen 2 and Gen 3 learned their behaviours entirely from their digital compatriots. No CPU and no Amiibo moves like a human does, and they can't do the things that a human can do. I would theorize that Gen 2 and Gen 3, in learning from Amiibo and not a human player, were stunted in their capabilities and played at a reduced level.

This would seem to contradict the findings of the experiment, though; Gen 2 surpassed Gen 1 in its Best-of-Five matches, as did Gen 3 to Gen 2. However, one must consider that the New-Gen Amiibo, trained from lv1 by a human player, then went and demolished very handily the Gen 3 Amiibo... arguably the most powerful and capable of the experiment until that point.

The data we can extrapolate from this is very simple: an Amiibo will learn well from another Amiibo, but specific training that comes from a human player will trump this. Even when it comes into contact with other Amiibo and begins to take on untrained characteristics, a human-trained Amiibo will retain enough of its training to have a serious advantage over its fellows.



Even with this data, I'm not sure I'm quite finished yet. As it stands, I'm tempted to extend the Best-of-Five challenge and take it to a more extreme number, like fifteen. This will give the Amiibo a chance to normalize, and see if the teaching I did can be swept away by simple exposure to an inferior Amiibo.

That said, I would be thrilled to hear any thoughts or theories of your own! Let me know what you think, and what experiments you've run or would like to see run.

Until next time, be awesome dudes and dudettes and get Smashing!

 

[UltimateXsniper]

I did a similar training session with my 2 amiibos.
Bidizzle my first amiibo. Currently the best.
Bozer. Trained by Bidizzle only.
Bozer had 3 other brothers as well.
Curbie, Sparky, Danky Kang. I have the full story on the general amiibo thread so I'll TRY to make it short.
, Trained by only level 9 cpu kirby.
, Trained 100% by me.
, Tries to handle 1v7s.

Before they start their training I set the 4 on a free for all to who wins. Bozer was last, Curbie 3rd, Sparky 2nd, Danky Kang 1st.

Once the training was over they did another free for all. Bozer won, Curbie 2nd, Sparky 3rd, Danky Kang last. However, I also did 1v1s. Sparky won all his matches against his brothers on 1v1s so he was the best of the 4 on 1v1s.

But enough about that, Let's focus on Bozer, the one trained by an amiibo Bidizzle. He never defeated Bidizzle once during training and still to this day. None of the 4 can defeat Bidizzle anyway. However I did a 1v1 against Bidizzle and he played a little differently. I could defeat hin easier than usual. It's probably because he had to train a weaker opponent for some time and had to learn about Bozer before Bozer can get the upper hand.

But I don't see any true improvement from the 2nd gen Bozer. And also made Bidizzle weaker in the process. Bidizzle is doing fine now though (No need to reset amiibos unless they are unbelievably weak). However, I was surprised that Bozer won the free for all. Consider he was last the first time they played, shows some huge improvement. You can tell he tries to space a lot which is probably why he won free for all but lost to Sparky on 1v1.

If you want to hear more about the 4 amiibos check out part 1 here http://smashboards.com/threads/amiibo-general-thread.384590/ It contains most of the information of their sessions while other parts are just the amiibos trying to get better from here on out. You might want to dig in that thread if you want to see the other parts. (There's like 4 parts) It's not very detailed compared to your theory but you might get some important notes on these different training session they went through.

 

[Kornaki]

This is not surprising at all. If you played against only a single player, would you become amazing over time? No, you'd suck because you would just get used to that player's tendencies, and nothing else. The phrase "learning algorithm" doesn't imply some magical artificial consciousness that will always elevate itself. I don't know much about how game AIs work, but I do know a good amount about how learning algorithms work. The basic idea of how the learning component works is that the character is controlled by several numerical parameters. When it's ten feet away from you, it decides to move towards you or roll away, and that's decided by a parameter that we'll call aggressiveness. If aggressiveness is .6, then there's a 60% chance it moves towards you, and a 40% chance it rolls away. When it attempts a smash attack, that's controlled by another parameter that well call smash factor. If smash factor is .7, then the character attempts a smash attack 70% of the time, and a tilt attack 30% of the time.

OK, so how does the character learn? Let's say it gets into a fight with you, and during the fight it runs at you twelve times, and rolls away six times. Those twelve times it runs at you, it hits you twice with dash attacks, misses a couple grabs that you spot dodge, and just gets smash attacked the rest of the time. Those six times it rolls away, three times you attempt to run at it with a dash attack and miss, and it follows up with a smash attack of its own that hits. The other three times you don't chase and nothing happens.

The game comes away from that deciding that rolling away has better outcomes than running towards you on average, so it is going to roll away more often. It decreases aggressiveness from .6 to .5. Similarly, it also saw that every time it smash attacked, it hit you and dealt a lot of damage! So smash factor goes up from .7 to .8.

This is in secret an example of an optimization problem. A very simple example of how this kind of process goes awry can be seen in single dimension function optimization. I can't post a link, but on wolfram alpha draw the graph y=(x-1)(x-2)(x-3)(x-6) for x between 0 and 7 (full syntax is "plot y=(x-1)(x-2)(x-3)(x-6) on x=0..7")

Suppose you are trying to minimize the graphed function (find the smallest y value, and the corresponding x value). One way you might do it is to plug in a value, say x=1. Then try x=1.01 and x=.99. You see that x=1.01 gives a smaller y value, so you move to the right, say to x=1.2. You try x-1.19 and x=1.21, and find that it is still going down as x gets bigger, so you move to x=1.4. You try x=1.39 and x=1.41, and find that you need to move to the left this time. You don't want to go back to x=1.2, so you go back to x=1.3 instead. Repeating this, you eventually get to x=1.399 which is a local minimum of the function.

Great, you found the minimum! Of course, anyone looking at the whole picture can tell you that your minimum sucks, and there's one at about x=5 that makes y a lot smaller. But if you just take small steps like I described, you will never reach that point. You need to make a drastic change.

The analogy with amiibo is the amiibo needs to see something drastically different. Throwing different characters at the amiibo is OK, trying different tactics is OK, but that's not as extreme as you could do. I just ran my amiibos on a 4v4 game and a 2v2v2v2 game to mix up the kind of scenarios they see. My yoshi, which has probably got ten hours under its belt already, got another hour in and learned to short hop into a timed sex kick out of a shell. I haven't seen him do that ever before, because previously he was stuck in power shield-> up smash. That's a fairly effective move in one v one battles, but is not as good in multiplayer I assume because of the lack of surrounding hitboxes.

If I had just kept fighting in 1v1 battles, and half the time he does the shield to up smash he gets a good hit, one quarter of the time he gets a bad hit or misses, and one quarter of the time he gets punished, then he will never learn to stop doing it, because on average it's still doing really good. But when he was put in a scenario where two thirds of the time some random guy comes over and destroys him, all of a sudden the value of that move went down A LOT. In 1v1 matches he could never learn to make the switch, because in those matches he was stuck in an area that was silar to the x=1.399 in the graph I showed above; small changes weren't going to get him to change his ways, he needed a big jump to do it.

The last thing, to talk about edge guarding. The amiibo is making small changes in how it fights, and there are a lot of different changesit can make. If it uses shield->up smash 95% of the time, and it does that ten times effectively during the game, and once during the game it sees someone try to edge guard and do it successfully, it still has ten times as much data saying "do shield to up smash". So when it makes the small step it will still put more weight into the shield->up smash (maybe deciding to do it 97% of the time) than it will into "go out and edge guard". So if you want your amiibo to learn more about being over open air, you need to force that scenario A LOT, rather than just throwing out a couple spikes in a three stock match.

 

[Faora Meridian]

This is not surprising at all. If you played against only a single player, would you become amazing over time? No, you'd suck because you would just get used to that player's tendencies, and nothing else. The phrase "learning algorithm" doesn't imply some magical artificial consciousness that will always elevate itself. I don't know much about how game AIs work, but I do know a good amount about how learning algorithms work. The basic idea of how the learning component works is that the character is controlled by several numerical parameters. When it's ten feet away from you, it decides to move towards you or roll away, and that's decided by a parameter that we'll call aggressiveness. If aggressiveness is .6, then there's a 60% chance it moves towards you, and a 40% chance it rolls away. When it attempts a smash attack, that's controlled by another parameter that well call smash factor. If smash factor is .7, then the character attempts a smash attack 70% of the time, and a tilt attack 30% of the time.

OK, so how does the character learn? Let's say it gets into a fight with you, and during the fight it runs at you twelve times, and rolls away six times. Those twelve times it runs at you, it hits you twice with dash attacks, misses a couple grabs that you spot dodge, and just gets smash attacked the rest of the time. Those six times it rolls away, three times you attempt to run at it with a dash attack and miss, and it follows up with a smash attack of its own that hits. The other three times you don't chase and nothing happens.

The game comes away from that deciding that rolling away has better outcomes than running towards you on average, so it is going to roll away more often. It decreases aggressiveness from .6 to .5. Similarly, it also saw that every time it smash attacked, it hit you and dealt a lot of damage! So smash factor goes up from .7 to .8.

This is in secret an example of an optimization problem. A very simple example of how this kind of process goes awry can be seen in single dimension function optimization. I can't post a link, but on wolfram alpha draw the graph y=(x-1)(x-2)(x-3)(x-6) for x between 0 and 7 (full syntax is "plot y=(x-1)(x-2)(x-3)(x-6) on x=0..7")

Suppose you are trying to minimize the graphed function (find the smallest y value, and the corresponding x value). One way you might do it is to plug in a value, say x=1. Then try x=1.01 and x=.99. You see that x=1.01 gives a smaller y value, so you move to the right, say to x=1.2. You try x-1.19 and x=1.21, and find that it is still going down as x gets bigger, so you move to x=1.4. You try x=1.39 and x=1.41, and find that you need to move to the left this time. You don't want to go back to x=1.2, so you go back to x=1.3 instead. Repeating this, you eventually get to x=1.399 which is a local minimum of the function.

Great, you found the minimum! Of course, anyone looking at the whole picture can tell you that your minimum sucks, and there's one at about x=5 that makes y a lot smaller. But if you just take small steps like I described, you will never reach that point. You need to make a drastic change.

The analogy with amiibo is the amiibo needs to see something drastically different. Throwing different characters at the amiibo is OK, trying different tactics is OK, but that's not as extreme as you could do. I just ran my amiibos on a 4v4 game and a 2v2v2v2 game to mix up the kind of scenarios they see. My yoshi, which has probably got ten hours under its belt already, got another hour in and learned to short hop into a timed sex kick out of a shell. I haven't seen him do that ever before, because previously he was stuck in power shield-> up smash. That's a fairly effective move in one v one battles, but is not as good in multiplayer I assume because of the lack of surrounding hitboxes.

If I had just kept fighting in 1v1 battles, and half the time he does the shield to up smash he gets a good hit, one quarter of the time he gets a bad hit or misses, and one quarter of the time he gets punished, then he will never learn to stop doing it, because on average it's still doing really good. But when he was put in a scenario where two thirds of the time some random guy comes over and destroys him, all of a sudden the value of that move went down A LOT. In 1v1 matches he could never learn to make the switch, because in those matches he was stuck in an area that was silar to the x=1.399 in the graph I showed above; small changes weren't going to get him to change his ways, he needed a big jump to do it.

The last thing, to talk about edge guarding. The amiibo is making small changes in how it fights, and there are a lot of different changesit can make. If it uses shield->up smash 95% of the time, and it does that ten times effectively during the game, and once during the game it sees someone try to edge guard and do it successfully, it still has ten times as much data saying "do shield to up smash". So when it makes the small step it will still put more weight into the shield->up smash (maybe deciding to do it 97% of the time) than it will into "go out and edge guard". So if you want your amiibo to learn more about being over open air, you need to force that scenario A LOT, rather than just throwing out a couple spikes in a three stock match.

Firstly, I'd like to thank you for your well-thought out and appropriately verbose reply. Dialogue about this experiment helps keep me interested.

Secondly, while I see the point you're trying to get at (if I might summarize it as: "Diverse training is better than specialized, limited training."), that was not the purview of this experiment. The point of this experiment was to engage the Amiibo in a controlled environment, in controlled conditions, such as to extrapolate their ability to perfect particular behaviours. The goal was never to create a perfect Amiibo, merely an Amiibo that would, through learning from its forebears, perfect a particular style of combat.

Thirdly, you pointed out a crucial piece of information during your post regarding how extensive behavioural training is necessary in order to make large shifts in tactical decision making on the part of the Amiibo. This is why I hand-trained from lv1 to lv50 one of them (two, if you count the New-Gen one) and then had them train each other similarly. This allowed the Amiibo to have a greater length of time with which to train themselves into adaptational behaviour. While you correctly point out that larger-scale battles would allow the Amiibo to reinforce its patterns (or learn new ones) faster, that is outside the scope of the experiment and my finances.

Finally, the only way I could expand on the experiment and engage them in the way you describe (which inherently defeats the 1v1 key aspect of the experiment itself) would be to have a complete set of 8 Samus Amiibo. I however have no intention to purchase another six Samus Amiibo just for the purpose of maintaining that control element to the experiment, but as mentioned, it defeats the purpose of perfecting a 1v1 Amiibo. Yes, I could use a wide variety of Amiibo instead, but that too is not the point of the experiment; the goal was to be one character specific.

As a side note in answer to your initial sentences, the only person a friend plays Smash with at present on the Wii U is me. He's improving drastically through my teaching him, and he only plays against me. While I do test him every now and again with different characters, most often he's set up against my Samus with his ZSS. Human beings reason in different ways to strings of programming code (and are capable of extrapolating data that a computer at present cannot), so the argument that a player would simply be trash because they only face one opponent is, itself, trash. Human capabilities have no place in this experiment, unless they're my capability regarding teaching efficacy.

 

[Kornaki]

Firstly, I'd like to thank you for your well-thought out and appropriately verbose reply. Dialogue about this experiment helps keep me interested.

Secondly, while I see the point you're trying to get at (if I might summarize it as: "Diverse training is better than specialized, limited training."), that was not the purview of this experiment. The point of this experiment was to engage the Amiibo in a controlled environment, in controlled conditions, such as to extrapolate their ability to perfect particular behaviours. The goal was never to create a perfect Amiibo, merely an Amiibo that would, through learning from its forebears, perfect a particular style of combat.

Thirdly, you pointed out a crucial piece of information during your post regarding how extensive behavioural training is necessary in order to make large shifts in tactical decision making on the part of the Amiibo. This is why I hand-trained from lv1 to lv50 one of them (two, if you count the New-Gen one) and then had them train each other similarly. This allowed the Amiibo to have a greater length of time with which to train themselves into adaptational behaviour. While you correctly point out that larger-scale battles would allow the Amiibo to reinforce its patterns (or learn new ones) faster, that is outside the scope of the experiment and my finances.

Finally, the only way I could expand on the experiment and engage them in the way you describe (which inherently defeats the 1v1 key aspect of the experiment itself) would be to have a complete set of 8 Samus Amiibo. I however have no intention to purchase another six Samus Amiibo just for the purpose of maintaining that control element to the experiment, but as mentioned, it defeats the purpose of perfecting a 1v1 Amiibo. Yes, I could use a wide variety of Amiibo instead, but that too is not the point of the experiment; the goal was to be one character specific.

As a side note in answer to your initial sentences, the only person a friend plays Smash with at present on the Wii U is me. He's improving drastically through my teaching him, and he only plays against me. While I do test him every now and again with different characters, most often he's set up against my Samus with his ZSS. Human beings reason in different ways to strings of programming code (and are capable of extrapolating data that a computer at present cannot), so the argument that a player would simply be trash because they only face one opponent is, itself, trash. Human capabilities have no place in this experiment, unless they're my capability regarding teaching efficacy.

You're teaching him, but you're also on this board, so I assume you have watched smash videos, maybe played against some CPU 9s, or for glory, or you have played against other friends it sounds like. So you have a diverse experience set that you bring to the table that you can communicate to your friend. When the computer plays, it doesn't really have access to that data. Imagine two people who have never played video games before, who got super smash bros and played each other for a week, not looking up anything online. They then go to a tournament. Isn't this the biggest stereotype in the world? They think they're awesome, and then they end up sucking. This isn't limited to smash, anytime you have a small community of people isolated from the larger community, they aren't going to be as good at the activity they are doing. What the amiibo is going through is a lot closer to this scenario than the 'experienced teacher mentors a new smash player'.

To clarify my large scale battle experiment, I only have four amiibo, and I threw in four CPU 9s to complete the battle. I understand the idea of the experiment. To go back to the graph example, your first amiibo probably found a local minimum by the end of his training. When you throw him up against another amiibo, that amiibo is probably being attracted to the same local minimum (especially if the amiiibo learns from when the opponent's attacks are effective or not). Then the third amiibo gets sucked into the same local minimum. Yeah, they get a little bit closer over time, but they're never really going to improve. You need to put them in a scenario where the old moves absolutely suck to make them change their behavior for the better.

Making the situation is worse is the playtime involved. Your amiibo have been playing for what, four hours to get to level 50? As everyone agrees, amiibo learn worse than humans do. If they try to edgeguard and die, you can't say "that's OK, it's worth it once you figure it out", they just associate edge guarding with dying. So maybe they learn at half the rate of a real human playing you. So it's like giving someone a controller, and two hours later expecting them to have picked up on all sorts of crazy moves. Computer learning in general is effective because computers can process huge data sets incredibly quickly. The amiibo doesn't get to process fights any faster than you do.

My point isn't to criticize the experiment; I think it's pretty cool. But people seem to be disappointed by the learning capability of amiibo, and a large part I think stems from not understanding how computer learning works. I think it's cool that this was tried (and it was definitely worth trying!), but you shouldn't feel like this disproves that amiibo are capable of learning.

 

[TweetyPurd]

I found with my kirby and captain falcon that if i constantly went offstage to edgegaurd, they learned to do it every time too!

However, i played in a 1v1v1v1, and just played it super safe, and ran away, and within a few games my amiibo would literally just run away from the opponents, barely attacking at all.

The problem with amiibo i find is that because they pick up on the opponents tendancys, if they play against someone who does somthing you dont want them to do, they will pick it up anyway

 

[Andy Summer]

"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."

Interesting theory. Very well thought through, but I must point out a flaw to this theory. As the player trains the amiibo, it learns to counter that specific player's style and character of choice. Throw something new at it, something that doesn't fight like the trainer, and the amiibo will struggle. The amiibo is powerful, but not invincible. My Captain Falcon can best me playing as Falcon, sure, but when I play Samus or Megaman, he struggles in the fight. Each character has their weaknesses that can be exploited at any skill level.