Training your amiibo: Don’t waste your time

[TeddyBearYoshi]

so is the OP wrong?

That's debatable.

-snip-

This doesn't feel all that relevant. We're talking about if the amiibo actually learn or are influenced by you, not whether the functionality of them is all that exciting or not.

 

[HeavyMetalSonic]

I feel like they can pick up certain traits, but not full blown combo's.

My Pikachu for example... I taught it to use quick attack to keep applying pressure to opponents, and now every time it gets to a certain range it'll quick attack towards them and back again. I also taught it to perfect shield ALOT, my friend got three stocked by it because he simply couldn't touch him. I also taught it to use Usmash as a kill move. Basic things like that it can seem to pick up on, much like Brawl's AI did. It's adaptive but not perfect (I taught Falco on Brawl to short hop into Dair, but he couldn't chain grab).

I don't believe it can learn massive combos though. I would try as often as I can to do several Uilts into Uairs on it, but it would never pick up on what I was doing, although it DID start using Utilts more. It also started using Usmash more as I tried to teach it to use it as a kill move. I tried to teach it down throw into Thunder, but it could never seem to do it. Instead it'd just try to use Thunder when anyone was above it or within grabbing range.

I think people THINK it's learning what it's doing because it's reacting to the situation. An example would be Kirby's Dair into grab or whatever. You teach Kirby to use it's Dair when it's just above the opponent. So it will keep doing that. But you also teach Kirby to grab when he's within a certain range of the opponent, and after the Dair it's in the perfect spot to do so, and being as their reactions are insane, it never misses. So it's not that he's learned to Dair into a throw combo, it's that he's learned to Dair in a certain spot, and he's learned to grab in a certain spot, so he's learned two things but one sets him up for the other. If you understand what I'm saying...

It's fairly obvious that they use the same AI as the CPU. Sonic is a great example of this. I can charge his SD in a second through tapping SO fast (I call it "vibrate tapping"), and I've noticed that whenever a Sonic CPU is present, the charge sounds are EXACTLY like the speed when I do it, because it's learned it off me.

I'll probably try out a few more things on different amiibos (Mario, Link, Fox, Falcon, Marth, Diddy, and when they come Ike, Bowser, Lucario, Toon Link, Sonic and hopefully Rosalina if I can get hold of one...). I'll let you know how it goes.

 

[Gidy]

That's debatable.


This doesn't feel all that relevant. We're talking about if the amiibo actually learn or are influenced by you, not whether the functionality of them is all that exciting or not.

So the answer is still undefined?

 

[TeddyBearYoshi]

So the answer is still undefined?

Yes, personally, I think he's wrong, but some people agree with him, some don't.

 

[Victor Marczyk]

I beg to differ, my bud told me that fighting my pikachu amiibo was almost identical to fighting me. Although amiibos tend to represent a bot more or less, they still adapt to your play style based on my experience with them

 

[stancosmos]

I have trained some of my amiibos again and again to learn how to trigger a behavior I want. I have a Yoshi resetted and trained more than 15 times and here is what I have learned:

amiibos do not imitate your Playstyle. They do not reproduce your moves or combos. The way you play does not have a direct impact on the behavior of an amiibo. It does passively influence it, as you will see later. amiibos most probably have some sort of „digital evolution“ I will explain later.
amiibos do not store anything close to complex AI. When you google, you find that there are 256 byte a game can store on a Figure besides Name, Owner etc. Thats not 256 Kilobyte, it’s 256 byte. That’s 256 numbers between 0 and 255. A single Windows Icon easily is more than 100 times that size. You can see that there is no room to store any complex behavior. But playing with an amiibo DOES result in very different play styles, so the DO have to store some sort of behavior, don’t they? Well, yes an no.



amiibos do in fact just use the AI all CPU-Opponents do use also. This AI has fixed behaviours for given „Moves“. A Move becomes „valid“, when a certain situation is „true“. Example: A CPU has a Side-Special with a Range of 80 Pixels. An Opponent 81 Pixels away will be missed, anything closer than 80 Pixels will result in a hit, but with decreasing damage done. For a Level 1 CPU this move will become „Valid“ when an Opponent comes close to anything between 100 and 50 Pixels. If the move becomes valid, the AI may choose to use it. It may choose which valid move it uses randomly or it has a priority list, I do not know. A Level 1 CPU will use this move with some chance to miss. For a Level 9 CPU this Move becomes valid then an Opponent comes close to anything between 83 and 50… as a result a Level 9 CPU will hit more precisely. This of course is just an example and much less complex as it really is, but this will be the way it works.



amiibos use the very same rules to fight. It only does not choose their moves randomly or by a fixed list, it does so by a changing priority. Here is how it works:



A Level 1 amiibo starts with random parameters for its Moves. It does not have any more Moves oder Combos than a Standard CPU has, it just has other priorities. This is why any resetted amiboo does seem to fight in another way than before. Lets again take the Side-Special as example. It gets an Priority-Value of 5. Then there is another Move that has a Priority-Value of 5 also. This is all that the amiibo stores. Now the AI of the game comes to play. Both moves get valid, when an Opponent comes close to a certain Range. As both have the same Priority-Value, they have a 50% Chance to be triggered. The Match ends with a certain result that is stored on the amiibo (lets say something like 0-100% success). The next fight the digital evolution kicks in: Priority Values are changed randomly. In our example the first Value is raised by one, the second lowered by one. The amiibo fights its match. It now uses the first move more often than the second one. After the match it stores its success-value. If it was more successful than last time, it will keep its new values. If it wasn’t, the changes are discarded. This way the behavior of an amiibo slowly changes to something it is successful with.



Additionally it gets more precise in its moves with a higher level, as it adapts the given rules of the moves as described earlier. And it also gets a damage boost, which makes High Level amiibos seem so much better than Low Level amiibos. You will find a Level 1 amiibo dealing only 50% of the damage that character normally deals with a certain move, while a Level 50 amiibo easily deals up to 150% of the damage.



If you play around with you amiibo, you will find it does not play as you do. It does however adapt to your weaknesses. Try it: If an amiibo alters its behavior in a way you like, let it pummel you like there is no tomorrow. It will keep the change. But in the end you cannot REALLY influence their behavior. The strongest and most aggressive amiibos I have I did not play against once while training. I just did let them Level by 8vs8 Smashes with amiibos only. That almost always results in 3-4 really strong amiibos and another 4 that are total crap. The good ones started to evolve in a good way, the bad ones did not find the right modifications… and they never catch up.



To get a good amiibo, it in not a good idea to fight against only one type of enemy. I tried. If you do, the amiibo will adapt to that exact enemy, but will be weak against others. If this exact enemy is vulnerable to a certain move, digital evolution will make this one move highest priority and other will almost never be used. You then have a specialist that takes a lot of time to adapt to new situations.



Of course all this is PROBABLY so, but it seems to be. And of course it is much more complex, as there are also rules involved how to react on moving platforms and things like that and then combinations of all those rules. But it can be taken as a fact, that there is only ONE Artificial Intelligence in the game that amiibos and CPU share, that you cannot train it anything this Artificial Intelligence does not already has in itself (like complete new Combos and such) and that trying to produce a certain play style for your amiibo will not work and is a waste of time. All you can do is tweaking the AI already there by letting it be successful or not.

My guess is that the reason you might be able to teach an amiibo while playing as the same character is it would judge the hit %s of both players after the match. For example if you used Yoshi's upsmash and hit the opponent 100% of the time, and he never used up smash, than he would still read your hit % and raise the priority of the move. I think it's just as simple as a value for hit% for each of its moves, and then a priority based exactly on that percentage.

 

[ChikoLad]

It's fairly obvious that they use the same AI as the CPU. Sonic is a great example of this. I can charge his SD in a second through tapping SO fast (I call it "vibrate tapping"), and I've noticed that whenever a Sonic CPU is present, the charge sounds are EXACTLY like the speed when I do it, because it's learned it off me.

Sonic CPU's aren't learning from your "vibrate tapping". Level 9 CPUs have an ability to "mash" buttons beyond what a human can. This is because they are not actually pressing buttons, but are just having inputs read out to them by the system.

On Level 9, inputs from the CPU are pretty much instantaneous. When it comes to commands that involve mashing a button repeatedly, like Sonic charging a Spin Dash, or a character trying to break out of a grab, their mashing is impossibly fast because it doesn't rely on human reflexes, but processing power instead, and the consoles running the game have enough to read out inputs in instantaneous succession. It was the same in all past Smash games.

CPU's also aren't like Amiibos, who retain the info they learn between matches. CPU's make small adaptations on a per battle basis, but still not near to the extent Amiibos can. Each character is programmed to have a specific playstyle when CPU controlled, with varying results.

For example, normal CPU's don't aim to attack Rosalina's Luma, nor do they register him as a threat. They only register his immediate attacks as threats, and will dodge or shield them, but they don't attack Luma back. This is because they haven't been programmed to target Luma like a player character, but they have been programmed to avoid incoming attacks, and Luma's attacks count as that, so they simply try to dodge or shield his attacks, without retaliating. This makes Rosalina completely broken against Level 9 CPU's in 1v1.

My Amiibos, however, have learned to attack Luma, and will do so if I leave him separated from me for too long, or if I get knocked away from my Luma.

 

[cwjakesteel]

The best use for Amiibo currently is decorating your room.

None of the digital functionality thus far is too exciting, although I do feel like Codename STEAM does it best. The actual functionality with Smash Bros. 4 doesn't excite me much, from a solo play perspective. I think, like everything else Smash, if it isn't enjoyed with other people, it's a waste. That's why I think pitting Amiibo against each other, with other people, is the only decent way to have fun with them in Smash 4.

Agreed. The most fun I have out of amiibos is just having them. Like figurines you put up for display. And tbh, I don't even know why I am so concerned about training my amiibo correctly, but my level 50 Wii Fit Trainer amiibo doesn't freaking mash B when doing her recovery and can never recover!!

If I wanted a good fight I would invite friends over or go play online. So what's the point of even having a "good" amiibo?

 

[digiholic]

Guys, 256 Bytes is more than enough to store complex artificial intelligence. This is because Amiibos are Genetic Algorithms. If you know about genetic algorithms, forgive me, because I'm way oversimplifying things here.

Genetic Algorithms are, effectively, a program that runs once, manipulates a few values, then runs again with the new values. Basically, things are given "weight", which is used to mathematically determine the course of action. If the A.I. runs and something doesn't work, it's weight goes down. If it does work, it's weight goes up.

Here is an example genetic algorithm. Cars are randomly generated and run a track. Future generations will be more like the cars that get further, meaning the code "evolves", just as life does, by passing on its "genes", or, in this case, weight.

An Amiibo doesn't need to store it's moveset, or any calculations, it just needs to store it's genes. The A.I. is already on the disc, but the values passed to it are given to the Amiibo.

Now, for a bit of math. A floating point number is 4 bytes long. A single precision floating point number can store a number between 1.175494351 * 10^-38 and 3.40282347 * 10^+38. This isn't an even distribution, as you start to lose some precision for very small and very high numbers, but what it comes down to is, a single gene can have 18,437,736,874,454,810,623 different values. Our Amiibo can hold 64 of these floating point numbers, meaning that the total number of distinct behaviors an Amiibo can have is 1.1800152 * 10^21. That's 1,180,015,200,000,000,000,000. That's One point one-eight sextillion different Amiibos. That's roughly equivalent to the number of stars in the observable universe.

You can do a lot with 256 Bytes.

 

[warriorman222]

So, to all of you claiming you have taught your amiibo something, are you sure that's not just confirmation bias? If they "learned" your combo, are you sure that's not just because it's a good one that the AI would do anyway? You can't really provide much evidence that what you did worked and wouldn't have just happened anyway.

On the other hand, while only 256 bytes of storable information on the Amiibo is pitiful, that doesn't preclude it from telling the FP to do certain things or not do certain things. That's enough bits to hold some information about general playstyles, but it would have to be interpreted console-side, and it would allow for only a teeny amount of variety, but it could still represent something.

Overall, though, I'm definitely siding with OP here. Unlikely that amiibos learn much.

OP is in deconfirmation bias. He refuses to believe that it can change anything at all. And there is lots of proof to prove him wrrng. Amiibos do learn stuff. Boundaries or no boundaries, they learn stuff. So OP is wrong.

 

[cwjakesteel]

OP is in deconfirmation bias. He refuses to believe that it can change anything at all. And there is lots of proof to prove him wrrng. Amiibos do learn stuff. Boundaries or no boundaries, they learn stuff. So OP is wrong.

Saying that an amiibo can learn is like saying that a submarine can swim.

 

[warriorman222]

Saying that an amiibo can learn is like saying that a submarine can swim.

I don't get it. Do submarines swim? Or do they not? 8th wonder of the world. Learn has 5 letters in it. 8-5=3.

I'm going to stop there before the army of paragraphs come.

Also this video proves that Amiibos learn from you. Yes both numbers could be closer, but chances are Thunderp is gonna consistently spam Thunder more than lvl 9 Pikachu.

 

[cwjakesteel]

I don't get it. Do submarines swim? Or do they not? 8th wonder of the world. Learn has 5 letters in it. 8-5=3.

I'm going to stop there before the army of paragraphs come.

It's referncing the original saying "Asking if an AI can think is like asking if a submarine can swim" (or something like that, don't quite remember it specifically). Anyway the point is that AIs and robots do what they are programmed to do. They do not have the sapience that grants them the qualities that only living things have like learning. Submarines move in the water, but they don't "swim."

 

[warriorman222]

It's referncing the original saying "Asking if an AI can think is like asking if a submarine can swim" (or something like that, don't quite remember it specifically). Anyway the point is that AIs and robots do what they are programmed to do. They do not have the sapience that grants them the qualities that only living things have like learning. Submarines move in the water, but they don't "swim."

Maybe they're programmed with Monkey see Monkey do? Just lie how all the tournaments do so with Apex? Explaining why I had to drop Miis? Maybe they can't learn form you, but they can adapt to you by copying you to an extent. Using the moves coded into them, and copying what you do. explains why when I trained my friend's Samus with Rosa, it would use dsmash just out of range when an enemy was in front, or using Nspec then throwing random moves out of nowhere (Luma was there so I could extend my range) The post below makes tons of sense.

Guys, 256 Bytes is more than enough to store complex artificial intelligence. This is because Amiibos are Genetic Algorithms. If you know about genetic algorithms, forgive me, because I'm way oversimplifying things here.

Genetic Algorithms are, effectively, a program that runs once, manipulates a few values, then runs again with the new values. Basically, things are given "weight", which is used to mathematically determine the course of action. If the A.I. runs and something doesn't work, it's weight goes down. If it does work, it's weight goes up.

Here is an example genetic algorithm. Cars are randomly generated and run a track. Future generations will be more like the cars that get further, meaning the code "evolves", just as life does, by passing on its "genes", or, in this case, weight.

An Amiibo doesn't need to store it's moveset, or any calculations, it just needs to store it's genes. The A.I. is already on the disc, but the values passed to it are given to the Amiibo.

Now, for a bit of math. A floating point number is 4 bytes long. A single precision floating point number can store a number between 1.175494351 * 10^-38 and 3.40282347 * 10^+38. This isn't an even distribution, as you start to lose some precision for very small and very high numbers, but what it comes down to is, a single gene can have 18,437,736,874,454,810,623 different values. Our Amiibo can hold 64 of these floating point numbers, meaning that the total number of distinct behaviors an Amiibo can have is 1.1800152 * 10^21. That's 1,180,015,200,000,000,000,000. That's One point one-eight sextillion different Amiibos. That's roughly equivalent to the number of stars in the observable universe.

You can do a lot with 256 Bytes.

 

[Nixon Corral]

It's referncing the original saying "Asking if an AI can think is like asking if a submarine can swim" (or something like that, don't quite remember it specifically). Anyway the point is that AIs and robots do what they are programmed to do. They do not have the sapience that grants them the qualities that only living things have like learning. Submarines move in the water, but they don't "swim."

As a computer science major with a focus in Artificial Intelligence, I can tell you that that is actually incorrect. There's a whole discipline focused on this! http://en.wikipedia.org/wiki/Machine_learning

Basically, automata and humans are not that different. If at all. We just haven't had enough time or research to make computers learn to nearly the degree that humans can, nor can we fully simulate neurons or anything like that (yet). Simulating intelligence and learning, though, are by no means impossible. If the organic matter of our brains can do it, so can the silicon and electrons of machines.

There's a certain conceit of humans that likes to believe that machines can't learn (to protect their own status as the highest learners), but it simply isn't true. Humans are nothing more than very complex machines that are very good at assembling and responding to new data!

 

[cwjakesteel]

As a computer science major with a focus in Artificial Intelligence, I can tell you that that is actually incorrect. There's a whole discipline focused on this! http://en.wikipedia.org/wiki/Machine_learning

Basically, automata and humans are not that different. If at all. We just haven't had enough time or research to make computers learn to nearly the degree that humans can, nor can we fully simulate neurons or anything like that (yet). Simulating intelligence and learning, though, are by no means impossible. If the organic matter of our brains can do it, so can the silicon and electrons of machines.

There's a certain conceit of humans that likes to believe that machines can't learn (to protect their own status as the highest learners), but it simply isn't true. Humans are nothing more than very complex machines that are very good at assembling and responding to new data!

Yea automata and humans are not that different...if you ignore about 50% of what makes us human...It's easy to simply the human being into a biological machine, but a machine has no inherent will. Anyway, the reason AI can't think isn't just because of a limitation of technology, nor because of a limitation of achievable intelligence but because an AI can only simulate those things. There's a reason it's called artificial Intelligence, because what seems to be a thing, is not really that thing. Anyway, this is off topic. I don't believe that automata not being able to think is a timeless truth.

 

[platomaker]

I've retrained my Mario 4 times now and it still won't use FLUDD lol

I never retrained my mario amiibo, just let it do its thing and eventually (after 3 months at level 50) it started using its FLUDD, only a little bit and ONLY when I use my main. Its hard enough to get it to cape, but its definitely getting there, be patient with it.

 

[Nixon Corral]

Yea automata and humans are not that different...if you ignore about 50% of what makes us human...It's easy to simply the human being into a biological machine, but a machine has no inherent will.

But "will" is just an emergent property of an extremely complex assembly of neurons! Anything we will ourselves to do is still the result of a response to the myriad stimuli we are constantly barraged with (or to satisfy our reward pathways, etc). Machines are capable of that too, More and more as time goes on.

Your arguments are super semantic. If artificial just means non-organic, it's meaningless to distinguish the two. There's very little meaningful difference, as a simulation approaches perfection, between simulating something and "actually" doing it. In fact, if a simulation reaches perfection, it's the same as actually doing it.

Anyway, this is off topic. I don't believe that automata not being able to think is a timeless truth.

It's not even true now! :D

But I don't think it's that off-topic. This thread is pretty inherently related to AI, so discussion thereof enriches it.

 

[cwjakesteel]

I never retrained my mario amiibo, just let it do its thing and eventually (after 3 months at level 50) it started using its FLUDD, only a little bit and ONLY when I use my main. Its hard enough to get it to cape, but its definitely getting there, be patient with it.

So amiibos still learn after they've reached level 50? That's a relief. The only problem is none of my amiibos are Villager, and I like fighting them with Villager, so does that mean they will (wrongly) try to imitate me while I'm Villager?

 

[platomaker]

So amiibos still learn after they've reached level 50? That's a relief. The only problem is none of my amiibos are Villager, and I like fighting them with Villager, so does that mean they will (wrongly) try to imitate me while I'm Villager?

I tried retraining my Wii Fit Trainer 3 times, she still REFUSES to mash B to recover so her UpB goes half the distance and she gives up halfway through that. I'm going to give it more time. The Mario Amiibo I have is seriously skilled, and that was just me dinking around. I have a Diddy and Yoshi that are doing stupid stuff on account of having them train together.

NOTE: do not train them with CPUs or against other Amiibos unless you want some discombobulated play.

 

[cwjakesteel]

But "will" is just an emergent property of an extremely complex assembly of neurons! Anything we will ourselves to do is still the result of a response to the myriad stimuli we are constantly barraged with (or to satisfy our reward pathways, etc). Machines are capable of that too, More and more as time goes on.

Will being an emergent property of our neurons is an unproven interpretation of the human mind. If you're a materialist of course AI can have intelligence equal to ours, but in the end it just means our intelligence is just as much a façade if we are at our highest, machines.

Your arguments are super semantic. If artificial just means non-organic, it's meaningless to distinguish the two. There's very little meaningful difference, as a simulation approaches perfection, between simulating something and "actually" doing it. In fact, if a simulation reaches perfection, it's the same as actually doing it.

Artifical doesn't mean non-organic...it means non-genuine. By definition, artificial intelligence is not intelligence. Yes that's semantic but the meaning of words matter because we use them....

It's not even true now! :D

But I don't think it's that off-topic. This thread is pretty inherently related to AI, so discussion thereof enriches it.

About Amiibo AI, I just want my Wii Fit Trainer to mash B when recovering!

 

[Veigar]

I have a Yoshi amiibo as well. The first time I trained it, it ended up always spamming the egg throw and egg lay moves at level 50 despite me never doing this as I was training it. The second time I trained it, the same seems to be happening.

 

[cwjakesteel]

I tried retraining my Wii Fit Trainer 3 times, she still REFUSES to mash B to recover so her UpB goes half the distance and she gives up halfway through that. I'm going to give it more time. The Mario Amiibo I have is seriously skilled, and that was just me dinking around. I have a Diddy and Yoshi that are doing stupid stuff on account of having them train together.

NOTE: do not train them with CPUs or against other Amiibos unless you want some discombobulated play.

But what if I want to take my amiibo with me and battle it against 3 of my friends' to see who has the best, and we do that for a few days. Will all of them become dimwitted from our tournament as a result?

 

[platomaker]

I have a Yoshi amiibo as well. The first time I trained it, it ended up always spamming the egg throw and egg lay moves at level 50 despite me never doing this as I was training it. The second time I trained it, the same seems to be happening.

Punish it brutally, make it cry digital tears and swear vengence. ...Then again my yoshi's a bonehead too so I'll have to get back with any progress. I say, don't reset and keep faith in it. Especially since you and I both know from experience that resetting it didn't help.

But what if I want to take my amiibo with me and battle it against 3 of my friends' to see who has the best, and we do that for a few days. Will all of them become dimwitted from our tournament as a result?

well you don't have to save the progress one, but if you want I don't really see a problem so long as its not exclusively trained- again, most of the changes that occur are small miniscule changes. I.E. Mario caping after three months at being level 50. You'll have to reinforce good and bad behavior...with your fists...in the game.

 

[cwjakesteel]

well you don't have to save the progress one, but if you want I don't really see a problem so long as its not exclusively trained- again, most of the changes that occur are small miniscule changes. I.E. Mario caping after three months at being level 50. You'll have to reinforce good and bad behavior...with your fists...in the game.

That is why I hoped that training would actually stop at level 50...It's a shame the game doesn't recognize any criteria for "improving" but only "do more of what that guy is doing."

 

[platomaker]

That is why I hoped that training would actually stop at level 50...It's a shame the game doesn't recognize any criteria for "improving" but only "do more of what that guy is doing."

What criteria would be use to measure the "improving" process? Since everything is situational I suppose this is the most practical way of doing things.

Recently I just started using a bowser amiibo, and oddly enough it started attacking, and eventually killing, me at way earlier levels than the other amiibos did. If it ain't broken, keep trying don't fix it, sure enough it started catching me in things. I'd much rather it continue learning so I can keep being "corrected".

I think they have they just mimic the computer and slowly emulate the player more and more. I wouldn't be surprised if they just use the CPU data on disc and simply tweak bits and pieces in order to save on memory space for each amiibo figure. Of course that last bit was just my conjecture.

 

[cwjakesteel]

What criteria would be use to measure the "improving" process? Since everything is situational I suppose this is the most practical way of doing things.

Recently I just started using a bowser amiibo, and oddly enough it started attacking, and eventually killing, me at way earlier levels than the other amiibos did. If it ain't broken, keep trying don't fix it, sure enough it started catching me in things. I'd much rather it continue learning so I can keep being "corrected".

I think they have they just mimic the computer and slowly emulate the player more and more. I wouldn't be surprised if they just use the CPU data on disc and simply tweak bits and pieces in order to save on memory space for each amiibo figure. Of course that last bit was just my conjecture.

Yes you're right. Without the amiibo practising our bad habits and being able to regress, we couldn't really say we trained them well because there would be no other way to train them.

So then, does that mean that when our amiibo gets a hit on us, it is actually getting worse because we didn't skilfully parry him? Since our amiibos are mimicking us, whenever they win, they actually start the next battle going easy on us?

 

[platomaker]

Yes you're right. Without the amiibo practising our bad habits and being able to regress, we couldn't really say we trained them well because there would be no other way to train them.

So then, does that mean that when our amiibo gets a hit on us, it is actually getting worse because we didn't skilfully parry him? Since our amiibos are mimicking us, whenever they win, they actually start the next battle going easy on us?

I think the process is gradual, but if they lose enough times you'll definitely notice a difference. And I think they are more concerned with any new situations- like if you force them off stage at low percents for the first time.

 

[Veigar]

Would it help if I were to put my Amiibo vs 3 lvl 9 CPUs that are teamed up against it? o-o

 

[platomaker]

Would it help if I were to put my Amiibo vs 3 lvl 9 CPUs that are teamed up against it? o-o

Try it out and report back your findings. Would like to see another perspective.
From experience it just makes the amiibo run away and campy. Often failing when it approaches.

 

[EcnoTheNeato]

From personal experience, and that of others, one thing that I haven't seen mentioned is this:

Amiibos seem to learn based off of a "reward" system. Moves that hit more often will be used more often (or unlearned if they miss too much). Also, moves that kill will be used more. I've done Thunderderp, too, because I thought it was funny. One way to get it to do that? Getting hit with Thunder, and then forcing myself to die because of it.

A friend of mine taught his Luigi to only down Taunt (including Taunt KOs on the edge) using that method.

So! Using movesets helps, as does simply having a certain "playstyle." But Amiibos, I've found, also learn to use moves by being "rewarded" with a KO by using them!

 

[KenboCalrissian]

I've taught my Kirby to dair > footstool an enemy off stage, and it does it pretty consistently. Forget teaching them how to deal with bosses, though.

I believe you about the Yoshi problems though. I had a Yoshi that was working almost perfectly along with my style between levels 18-24, and then it gradually gave in to spamming Egg Toss/Egg Roll/Lick. It's really weird, because Yoshi's my main and I never play the way this thing does.

I suspect what's going on is at some point - maybe around level 25? - it switches gears, from "How do I play like my trainer" to "Let's start carving out the moves I'm using that aren't resulting in victory." It stands to reason that if I keep winning, it gets "frustrated" in a sense and starts negating its own likelihood of using moves - after all, the strategy you taught it isn't resulting in victory, so it must try to do the things you're not doing, which in Yoshi's case is Spam & Eggs.

My theory, then, is that it's more effective to beat the tar out of it up until ~ 25, and then start losing to it on purpose. Let it hit you, and then just allow yourself to tumble into a combo or death. Theoretically, that will reinforce the behaviors you want and convince it to keep using them.

Unfortunately, if I'm right then it kinda means a well-trained amiibo is only good for a few sessions - after you start playing it for real, it may start degrading in performance with this strategy. I'm going to retrain my Yoshi and test this theory.

 

[Viriquin]

My Mario amiibo sometimes will follow up a down-throw with an Up-B or D-Tilt. Does anyone else's Mario amiibo do this?

 

[platomaker]

I've taught my Kirby to dair > footstool an enemy off stage, and it does it pretty consistently. Forget teaching them how to deal with bosses, though.

I believe you about the Yoshi problems though. I had a Yoshi that was working almost perfectly along with my style between levels 18-24, and then it gradually gave in to spamming Egg Toss/Egg Roll/Lick. It's really weird, because Yoshi's my main and I never play the way this thing does.

I suspect what's going on is at some point - maybe around level 25? - it switches gears, from "How do I play like my trainer" to "Let's start carving out the moves I'm using that aren't resulting in victory." It stands to reason that if I keep winning, it gets "frustrated" in a sense and starts negating its own likelihood of using moves - after all, the strategy you taught it isn't resulting in victory, so it must try to do the things you're not doing, which in Yoshi's case is Spam & Eggs.

My theory, then, is that it's more effective to beat the tar out of it up until ~ 25, and then start losing to it on purpose. Let it hit you, and then just allow yourself to tumble into a combo or death. Theoretically, that will reinforce the behaviors you want and convince it to keep using them.

Unfortunately, if I'm right then it kinda means a well-trained amiibo is only good for a few sessions - after you start playing it for real, it may start degrading in performance with this strategy. I'm going to retrain my Yoshi and test this theory.

If you lose to it on purpose then that defeats the purpose. Keep beating the crap out of it while using your own style and it'll adapt within months. Use a wide variety of characters. The idea is that its going to use whatever it can to hit you and if it's spam is not doing the job then it too will fade away after a while...unless it meets a new character.

Again, whenever it plays against a computer it ends up with a "runaway" mentality.

 

[KenboCalrissian]

Months?? No way I'm going to invest that kind of time in a Yoshi that does nothing that throw eggs in the off chance it might change its mind one day. Surely there's a more efficient way of doing this.

 

[platomaker]

Well, you improve by playing em, so I'm not saying you have to grind everyday, just eventually (a few months) you'll see a change just by playing casually against it. You could alternatively just do 3- 99 stock matches to force a change, but I don't know how you'll body will handle it.

 

[Kornaki]

Stick in a 4 way fight against 3 level 9 computers for an hour, then come back and play it again,. You should notice a change in itsfighting style. The good stuff should stick around after 5-6 1v1 fights, the bad stuff should start to go away. You don't need to train these things for a long period of time to see behavior change.

 

[Mewtwo929]

From personal experience, and that of others, one thing that I haven't seen mentioned is this:

Amiibos seem to learn based off of a "reward" system. Moves that hit more often will be used more often (or unlearned if they miss too much). Also, moves that kill will be used more. I've done Thunderderp, too, because I thought it was funny. One way to get it to do that? Getting hit with Thunder, and then forcing myself to die because of it.

A friend of mine taught his Luigi to only down Taunt (including Taunt KOs on the edge) using that method.

So! Using movesets helps, as does simply having a certain "playstyle." But Amiibos, I've found, also learn to use moves by being "rewarded" with a KO by using them!

I agree with this. I fought my Zelda using Kirby spamming grabs till around lvl 10. Then all I would do is down b and shield and whenever she grabbed me I wouldn't touch the controller until I fell down to my death or waited a couple seconds while I'm laying down.

 

[Mewtwo929]

Bout' around lvl 30 and higher she started spamming grabs like there was no tomorrow.

 

[platomaker]

Stick in a 4 way fight against 3 level 9 computers for an hour, then come back and play it again,. You should notice a change in itsfighting style. The good stuff should stick around after 5-6 1v1 fights, the bad stuff should start to go away. You don't need to train these things for a long period of time to see behavior change.

Hearing you say that, I don't know what to think. But whenever they fight CPUs they tend to go back to being campy and running away.