• Welcome to Smashboards, the world's largest Super Smash Brothers community! Over 250,000 Smash Bros. fans from around the world have come to discuss these great games in over 19 million posts!

    You are currently viewing our boards as a visitor. Click here to sign up right now and start on your path in the Smash community!

Another Amiibo Discussion : How to make interesting opponents

shockfrost

Smash Apprentice
Joined
Jun 29, 2002
Messages
154
Location
Oz
Current Version: 1.0

Welcome to my Amiibo research topic. This topic divides up into 4 core sections.

1. What We Know about Amiibo (And what we don't).

Amiibo draw their actions, their fighting style, from three sources of information.
Those three information sources are:

- INHERENT: Each Amiibo taps into a database of pre-programmed behavior that explains how to perform basic tasks. The largest component of Inherent data is Offstage Recovery. The second largest is Proper Item Usage (such as attacking a smash ball) and the third largest is Stage Element interaction...

This inherent information cannot be changed, but it is usually Low-Priority - an Amiibo can develop a habit that overrides its inherent approach.

For example, all Amiibo know, at level 1, how to efficiently use their jumps, Up+B, Side+B and/or Stage-Grapple to recover in an effective manner, but this behavior can be overriden - a move that is usually used for recovery, such as Shiek's BOMF, can be used to attack enemies trying to ledge guard instead. As another example, an Amiibo can learn to use a Fairy bottle offensively, throw it offstage, or hold onto the item.

All Amiibo begin with all their attacks, defenses and techniques already known entirely. An Amiibo begins by attempting a move at random, and if it succeeds, the move is weighted to be performed more often. This initial learning can be called Discovery, and it generally only occurs when the Amiibo is very young.


- CONDITIONED: Amiibo can learn by attempting a move and having it succeed or fail. If the attack succeeds, the attack is weighted to be used more often. If it succeeds especially well (Striking multiple foes or causing a KO), the weighting is changed accordingly. Amiibo can change their conditioning based on remaining time, own or enemy percent, location on the stage, distance to the enemy, and possibly other factors. This can include non-attack maneuvers.

- DEMONSTRATED: In addition to learning from attempting a move and succeeding, an Amiibo will learn from falling victim to an attack. If an enemy attack proves effective, a two-fold adjustment happens: A heavy negative weight is placed on the move the amiibo attempted (Conditioning), but also, a lesser positive weight is placed on the move the enemy used. In this way, Amiibo can adopt a player's unique attack patterns, such as Bowser short-hopping into Neutral+B, or Pikachu performing a fly-by aerial Down+B.

This process is slow, and teaching longer, complex combos can be very difficult. But if the Amiibo attempts his enemy's attack and succeeds, he will weight it heavier, and it will be used more often. Amiibo learn relatively little through Demonstration... past the first couple exposures, the Amiibo learns just how effective that technique is, via Conditioning. But the initial discovery via Demonstration is none the less a priceless learning mechanism.

WHAT WE DON'T KNOW

It is still disputed what is exactly learned at each level.

For its first outing, it appears the Amiibo possibly starts the round with a random selection of 6-8 of its core moves, and attempts those moves. This is the Amiibo's chance to develop its own candid fighting style. As that and other fights progress, the Amiibo learns via Discovery and Demonstration, but it only seems to process Conditioning at the end of battles - so at any time, it might bust out a new attack and give it a try, or it might copy something you do right away, a few times. But it only seems to significantly adjust its style between matches.


2. Known and Accepted Training Methods

The Organic Method: Simply play against the Amiibo in your preferred setting, with all your skill, and defeat it / work together with it until its skill level rises.

This method is largely limited by your own skill. Note that it will keep learning past level 50. The Organic method has the advantage of developing Counter-Intuitive Style - it will naturally look for ways to defeat you, and may happen across methods that you hadn't thought of, or identify weaknesses in your fighting style you weren't aware of. It also may borrow any tricks or techniques you use that it finds especially effective.

However, the drawbacks of the Organic method include the "Corner-Cutting methodology" - instead of learning a 'proper' fighting style, many Amiibo develop a style that exploits a flaw in your own fighting, and many organic Amiibo are thought of as excessively cheap - constantly using whatever seems to work best, whether that is shielding, dodging, aerials, grabbing, projectile spam, or just plain running away. You also get highly limited agency over your Amiibo's development - you're gonna get what you get, and hopefully, it will be strong enough to beat you. With that said, some Amiibo do not effectively develop when Organically trained, such as Sheik.


CPU-9: Just what it sounds like, you can have your Amiibo fight against Lv.9 CPUs. Doing this, the Amiibo will become quite addicted to dodging, grabs, and generally sloppy, uninspired fighting offset by incredible reflexes. The big advantage to CPU-9, is that your Amiibo will become adept at Close-Quarters Reflex. It will air-dodge and powershield with the best of them, to the point of trivializing your effort to kill it.

But your Amiibo may not develop any uniqueness to its fighting style, leading to an Amiibo that feels just like a Souped-Up CPU-9 with better reflexes and a meaner stare. This doesn't always happen - there's a chance that Conditioning will result in some interesting emergent styles that are capable of punishing CPU-9. The odds of this happening vary depending on the exact matches the amiibo was put in, as well as its type. But since you have no control over the match in CPU-9 training, whether you get an interesting Amiibo, or a CPU Carboncopy, is mostly up to luck.

Arhat's Engineered Approach: A measured, segemented approach to controlling an Amiibo's behavior by carefully controlling its conditioning and performing specific demonstrations.
http://smashboards.com/guides/comprehensive-amiibo-training-table.541/

Cloud9's Approach: Reddit's best guess. gj internets. At least it's quite recent,
https://www.reddit.com/r/amiibo/comments/3dbkl1/clouds_amiibo_training_guide_v16/
and a continuation here
https://docs.google.com/document/d/1v8MtMA0vKwbx-RiciG0CGZItwfOb4oVxAnP1fDFMPJE/

Amiibo Vs. Amiibo: An interesting method. This method increases growth speed, and Amiibo more rapidly steal each other's techniques - because they have high confidence in other Amiibo's ability to make good decisions, they weight Demonstration much more heavily - very differently trained Amiibo of the same character will copy each other's successful moves VERY quickly.
Of special note is Evrys' Amiibo Trainer Source, which specializes in Amiibo vs. Amiibo training.
http://amiibotrainersource.blogspot.com/



3. Experimental Systems, Possible Training Methods

Of note is The Amiibo Trainer Podcast, which delves deep into Amiibo theory and application. Some of his "discoveries" are disputable, but hey, even Benjamin Franklin thought plants lived on electricity at first. This guy is at the forefront of the Amiibo Training scene, check his work out.
http://gcravens.com/the-amiibo-trainer-podcast/
If you like his work you can check out amiibotrainer.com, but this podcast is quick and informative.


You can make an Amiibo fight a perfectly still Player 1. This interesting exercise allows the Amiibo to develop its own skill set seemingly at random, including both lethal and completely non-lethal ones. It never receives negative conditioning this way unless it randomly KOs itself (and THAT causes some REALLY weird behavior) see reply 1 for an example.


4. The Quirks of Individual FPs, and methods to overcome them/train them in special ways.

None to speak of yet.
 
Last edited:

shockfrost

Smash Apprentice
Joined
Jun 29, 2002
Messages
154
Location
Oz
This topic isn't done yet, but a quick post to add the interesting results of a decent experiment from last night.


PEACH vs. NOBODY, Lv.1-50

I trained an Amiibo(Peach) from 1-50 by holding still with a variety of characters in 2-3 minute Time matches, on Yoshi's Island 3DS.

The Amiibo started off the first round by randomly selecting a handful of moves and throwing them out, without any attention to the distance to the enemy. At one point it attempted Side+B, Peach Bomber, and knocked me off the stage. It then continued to putz around randomly until the round ended.

Round 2, Peach was spamming Peach Bomber almost nonstop. It took several stock off the player before time ran out. It then proceeded to gradually add new "attacks" to its repertoire... but some of those attacks were grossly ineffective. The Amiibo never UN-learned an attack that it tried. It would Peach-Bomber back and forth across the floor, while the recently KO'ed player was up on the platform, and just miss and miss. Then it would hop up and knock the player off.

At level 18, it picked up an odd habit when it knocked the player to the ledge, then accidentally ran off the ledge itself. Instinct kicked in and it quickly grabbed the ledge - and the player fell and died due to edge-hog. Somehow this highly complex behavior was registered as a SUCCESS. Peach Amiibo actually developed a habit of running and grabbing a ledge away from a ledge-clinging player, something like a "Ledge Hog attack".

Over time, it would try to ledge-hog a player near the edge, but not actually clinging to the ledge... it would also learn other ways to defend the ledge such as Down+Smash, or jumping off, and hitting the hanging player from below with her umbrella.

Peach continued to develop (bad?) habits such as rigorously Peach-Bombing off the side of the stage when the enemy was sent flying, then quickly recovering herself.... Using the Up+tilt attack randomly when well out of range... and then Taunting after ledgehogging the player to death, or when killing the player in a matter of seconds. When facing Kirby around lv.30, it Up+Smashed it to death, and from then on it would occasionally Up+Smash enemies to death rather than Peach-Bomber them to death. It took very little time to figure out how to charge Up+Smashes until the foe had drifted into striking distance. Around lv.40 it would begin to float off the stage above slower-falling players, as if trying to guarantee their end. It had developed various aerial attacks, but was especially fond of Down+Air (whether it hit or not), and never used Up+Air.

At the conclusion of the exercise, the Peach Amiibo retained every single habit it had randomly adopted during its "training".

It never pulled vegetables, never Dash attacked, only occasionally Side Smashed (usually at nothing), never grabbed, dodged, sidestepped or shielded, only used its umbrella for recovery, never tried to attack the player from above, and performed Peach Bomber for roughly 65% of all its attacks. It would occasionally Down+Smash a player, in particular on the smaller center platform, and sometimes would repeatedly Up+Smash a player until it died. It would Peach-Bomber back and forth the lower stage while ignoring the upper platform, with no visible effect. It would walk around on the lower platform performing Up+Tilt, Up+Smash and Side+Smash. And if the enemy player grabbed the ledge, it would do one of four things: Run off the ledge and then ledge-hog (And then taunt when the player died), jump off the ledge and umbrella the player from below, Down+Smash from the edge, or run to the far side of the platform and just watch the player fall. Its motion pattern was especially lackadaisy, it often walked (skipped) towards the player, and periodically would stop playing for 2-5 seconds.

From this test, I have some theories.
1. Amiibo only generate Negative Weighting when damaged or KOed. Missing is not enough to convince an Amiibo to stop using a move. If an Amiibo is never damaged and never KOed, it will never learn when it shouldn't use an attack, and will gleefully keep every attack it "learns".

2. Amiibo receive Positive weighting when inflicting damage or a KO. This does NOT only apply to attack inputs! It can also apply to highly complex control inputs! Peach Amiibo was able to develop a habit of Ledge-hogging as an attack... After grabbing the ledge, the player was quickly KOed, and the Amiibo made the connection. As a result, it was repeated! Then, when Ledge-Hogging had a 100% Kill Success ratio, it was adopted as a primary form of Edge Guarding!
 
Last edited:

shockfrost

Smash Apprentice
Joined
Jun 29, 2002
Messages
154
Location
Oz
OK, this can be at v1.0. I think there's enough information to be worth looking at now.
 
Top Bottom