EPM
Quiet Member
I seek battle, not badges. Pride, not prizes. Mastery, not victory.
Posts: 35
|
Post by EPM on Mar 10, 2016 18:40:24 GMT
amiibo are a frustrating yet addictive Black Box. Over time, many battles and trainers have provided a wealth of data, and I've spent much time trying to understand the amiibo brain. The specific information is impossible to know or reverse-engineer, and data-mining has yet to crack the encryption. There's a lot of speculation, and a vast wealth of specific training details both on this site and others, but what I've sought is the underlying 'rules' to all amiibo, in regards to Smash figure players. I will also provide a few of my own independent observations and training methods. This is a lot of information, so I'm not going to try to post all of it at once. This post will be part 1, and I'll add more with time. 1. The NFC Itself
First off, the actual NFC chip in the amiibo. Thankfully, homebrew efforts have analyzed it for me. Translating from jargon, the takeaway is admittedly minimal, as knowing how the NFC's data is partitioned is of little if any use to interpreting AI patterns. What I consider an important detail, however, is that the NFC stores a value for the number of writes to the chip. Every time data is saved onto the NFC chip, a secret number ticks upwards. Interestingly, the Mii Fighters must have a unique structure and may have an altered data storage, as they must store two different Miis: Themselves, and their owner. All amiibo store their owner's Mii data, as you already no doubt are aware of because the first thing any amiibo makes you do is register a nickname and that. The Mii data for the owner is a decent sized chunk of the NFC's relatively small space, so if a Mii Fighter has to store twice as much of that data, I suspect that means they're somehow adjusted to accommodate that. Further delving into the depths of unknown data reveals that Smash itself dos most of the "sanity checking" and ensures the amiibo data is in useful, valid condition. The important info is at the bottom: "Overwriting all but the first 0x4-bytes with 0x7F/0xFF via memory writes doesn't cause any smash-3ds v1.0.5 crashes at all. Doing the same except with random data and with actually writing to the Amiibo NFC tag, doesn't cause any crash either." Translation: You can overwrite an amiibo NFC with pure nonsensical garbage aside from the bits of data that validate "hey this is an amiibo!" to the 3DS. Smash will then straighten everything out into working order. Of additional interest is that the Figure Player's level is clamped to 50. You've likely seen people with hacked amiibo with 200/200/200 stats for example, and despite that being impossible to create in-game, Smash will still accept and use that data. The level, however, being clamped to 50 means that there's no "behind the scenes leveling" going on. amiibo benefit greatly from experience, but it doesn't take the form of "level 51-70" for example. But what of FP behaviors? Sadly, it's impossible to tell how the allotted storage ranges for data are used. However, the NFC chip lends clues, and a reasonable guess can be made. NFC chips are arranged in "pages" each consisting of 4 bytes, though I intend to avoid overly-technical discussion here. The important part for the research I'm presenting is that these are the ranges of data the amiibo NFC chip actually allows a game/player to write to. Due to the information being laid out in hex, I'm a bit disoriented about the exact ranges, but those numbers don't matter anyway. The breakdown of write-enabled space: Page 4: Used for data encryption as well as storing the "number of times the amiibo has been written to" value, which is an unsigned 16-bit integer. In other words, that value literally starts out like this in binary: 0000 0000 0000 0000. The highest number it can store is 65,535, and due to it being 'unsigned', this value would hypothetically roll over to 0 if anyone ever somehow wrote to their amiibo that astronomical number of times. Due to how binary works, the next-smallest available size for storing this kind of data would be 8 bits, which can only go up to 255. In the scale of actual real-life usage, I doubt any of us have made a drop in the bucket as far as write totals go. A given user's most-used amiibo would be maybe a few hundred writes tops, and with extensive time maybe they could break into the thousands. But that's still a very minimal amount compared to the potential total, which leads me to believe that the number of writes has no bearing on any sort of "experience" value. Experience matters, but is not simply tied to how often the amiibo figure has been accessed by the system. The much more humanly possible 8-bit integer would have been used if so, and it'd need some extra code space to prevent it from overflowing back to zero. Pages 5-8: Additional validation data. Essentially, every time you use your amiibo or make a change to it, it is re-encrypted based on the data, and presently only Nintendo's software can actually match this algorithm to, y'know, make use of the amiibo. Powersaves is peeling apart some of the less-protected areas, but are not at all able to fully "unlock the box". Amiiqo doesn't even bother, just being a storage case for the encrypted files that simply presents them to the system. Pages 20-28: Minus the first 3 bytes of page 20, this is the first truly important range of data. It stores the Application Data for the amiibo, and this is why amiibo used for Smash figure players can't be used to store data from Mario Party 10 or Animal Crossing, and vice-versa. There's only so much you can pack into a single NFC chip, and as each page is 4 bytes, this means that a total of 32 bytes are devoted to establishing the amiibo for use with Smash Bros. In addition to the AppData, it also stores the amiibo's settings. How this space is used varies between games, but as was linked before, this has been disassembled for Smash already. It contains the stuff the user sees, basically! Stats, equipment, which special moves are set, costume, and so on. Pages 28~73: Section 1 of the encrypted buffer. 114 bytes. Pages 6D~82: Section 2 of the encrypted buffer. 54 bytes. Considering that the NFC's size in total is 504 bytes of data (with more used for structure and validation), these two ranges are easily the most relevant to training amiibo. These are where the usage data and "memories" of an amiibo are kept. That they're divided into two sections is interesting enough. To give you some perspective, this post itself contains far more data than an amiibo NFC. Every letter is 8 bits of data (1 byte) in size, and then factor in things like hyperlinks, formatting, and spacing for user readability. A tweet of 140 characters is more than a quarter of everything an amiibo could possibly hold. My point is that with such little space, we can assume that the space is used as efficiently as possible, because otherwise the product wouldn't really function at all. Since we're all here this far, we can conclude that amiibo do in fact learn and adapt with time, meaning that the internals are well-used. Hopefully.
One such bit of deduction is the "loot" amiibo bring back to you after battles. Powersaves indirectly reveals how this works, because Powersaves can't give you specific drops. If you want to get, say, a Critical Hitter X Attack piece of equipment, you can't just select that. Powersaves just fills the amiibo with random stuff for you. But then, also factor in how much of a waste of precious, limited space it would take for amiibo to store specific prizes in their minimal capacity. You'd need every single trophy's ID number, custom move IDs, and equipment IDs. Even if refined and stored as simply as possible, you'd then have to store an unknown, varying amount of loot depending on how much the amiibo had been battling. It'd be complicated and wasteful, and with such limited data space, that's unacceptable, thus improbable. But then look back at Powersaves. It generates loot randomly a given number of times. The trick is revealed! The amiibo don't store loot at all, they store the amount of loot, and then simply let the game itself generate the results when the NFC chip is read! And that's much, much easier to store as a single number. So how is that useful to training? Well, for example, I used to worry that the amiibo would set aside memory for loot, which would inhibit learning if too much of it built up without being 'flushed out'. This turned out to be foolish of me, and I'm happy to have been wrong about that. You can let loot accumulate all you want, it won't affect learning or training at all. Speaking of, not all battles are equal. I've noticed a distinct trend that my amiibo get more loot from more 'intense' battles, and amiibo in their level 1-50 period get significantly more loot for short battles. Additionally, the "1 stock match and self-destruct" method of building up an appetite to keep feeding amiibo equipment does not generate loot. The game's own tips mention that amiibo learn more from "more intense" battles, and the consensus here in regards to amiibo training supports that. Since the AI is essentially a decision tree, think of amiibo learning as "do this more" and "do this less", at its most basic. An 'intense' battle is one that provides a wealth of positive and negative reinforcement, via hitting and being hit with attacks, and establishing cause-effect relationships of "doing this attack at this range is most likely to deal damage". Since loot amounts are likewise seemingly tied to battle intensity, my estimate is that we can use the amount of loot as a rough measure of how much our amiibo learned from a training session, for better or worse. This is only a guess, however, and actually correlating the two is going to take more analysis. At any rate, that was the first dump of my research, having focused mostly upon the NFC structure and what I could guess at based on it. The way the "amiibo memory" is split into two partitions of different size is very interesting, but I can't make any kind of estimate about that. I want to think that it corresponds to "short term" and "long term" memory, but that's wishful thinking of mine and in no way based upon evidence.
|
|
Z
Full Member
Now I am become $wag, the moistener of your mothers' granny panties
Posts: 223
|
Post by Z on Mar 10, 2016 18:55:48 GMT
Thanks for clearing up Loot Generation. So is a 4-man Coin Army 99-stock match NOT the best way to get loot, then? Sometimes one or more of my CA figs won't get any Parts, but will only have a ton of coins after a 99-stock farming session. Why does this happen? Does Smash's RNG just hate me? (Gimme that Lifesteal, & Imp. Esc, ★★★★★★★★★★★★! I NEED it!)
|
|
EPM
Quiet Member
I seek battle, not badges. Pride, not prizes. Mastery, not victory.
Posts: 35
|
Post by EPM on Mar 10, 2016 19:30:29 GMT
Nah, that's still the best way. You can't beat bulk volume, though if I had to guess, there's some diminishing returns for battles of that length. I've hit the max of 9999 coins on an amiibo once in a similar situation. When it comes to farming, I'd recommend a 4-man match set to 10 minutes, and then just repeat that over and over for an hour or so. Granted, I have the 3DS version, which makes doing that substantially easier. It depends on what's convenient for you.
That said, Smash Run and Classic 9.0, Crazy Orders, etc. all openly mention in the game tips that they give better rewards for better performance, and that's an entirely different process than relying on amiibo random loot generation. That Smash 3DS can be cheesed for equipment on the reward payout wheel before a match is useful (pressing the home button allows for input buffering to stop the wheel precisely), though that doesn't change the fact that Classic mode is infuriating on level 9. Especially how frequently the "2v2" match comes up, AKA 1v2 on account of the teammate being unbelievably stupid and suicidal without fail.
|
|
Elishima
Undertale Pass Holder
Let's-a-go!
Posts: 918
|
Post by Elishima on Mar 11, 2016 4:47:59 GMT
O____________________________________O
Holy moly my brain...
No seriously, you are one heck of an observant fellow. I'm impressed!
|
|
EPM
Quiet Member
I seek battle, not badges. Pride, not prizes. Mastery, not victory.
Posts: 35
|
Post by EPM on Mar 15, 2016 22:16:07 GMT
2. The Amiibo AII linked a previous post of mine from another thread in part 1. As much of what that says is relevant here, I'm going to repost it in whole, and then elaborate. I like to think of Amiibo decision making as a Plinko board: Experience and training is all about adjusting the "pegs" (that is, weightable factors into what decision the amiibo ultimately makes) on the board, or walling off certain trajectories. This is a decision-tree, basically, but easy to visualize. Experience is the ultimate deciding factor for amiibo competence, I'm certain. I wish I had a way to tell how many battles mine have fought, but lack such means. A "bigger board" of more experiences offers a greater diversity of considerations, and also arranges the pegs to favor some outcomes over others based on immediate circumstances.
An excellent example is Mewtwo's Shadow Ball. Spending the time to charge it, rather than peppering a foe with weaker shots to control the neutral game, is something one needs to teach the amiibo early on. Indeed, I did not use uncharged Shadow Balls once in the course of its 1-50 period. As Mewtwo grew exposed to other amiibo, however, it learned to space with uncharged Shadow Balls anyway, which led to a frustrating tug-of-war of influences on its habits. You see, an amiibo can't really plan ahead. Only computers much, much more powerful than the processing space allocated to the amiibo AI could actually extrapolate current conditions and then prepare accurately. So from an amiibo perspective, what advantage does charging Shadow Ball offer? The opponent is able to draw in closer, it leaves them immobilized and vulnerable, and even if shielding or dodging out of the charge is feasible, it still puts them in an unfavorable situation.And a charged shot is an all-your-eggs-in-one-basket attack. Over time, however, my mirror matches drilled in the idea of charging up Shadow Ball when at a safe distance (the minimum seems to be 3/4 of Final Destination's length), immediately upon scoring a KO, or via respawn invulnerability. The latter two are inconsistent, but happen increasingly often, so I can definitely say they're being picked up on. One very clear case of adaptation, though, is that my Mewtwo amiibo will almost always charge up a Shadow Ball when I myself do. This is because, obviously, I'm not a threat to it while I do so, and it's learned I have a habit of fully charging. This lead to mutually reinforced habits, as I would fully charge while it did. All of our mirror matches now start the same way unless I deliberate break pattern: both of us fully charging up a Shadow Ball. So ultimately, experience taught Mewtwo the idea of how far is 'safe distance' to charge SB, and if it's any closer, it will space the opponent with uncharged shots to zone them. A favorite trick I taught it was to pepper shots while slowly letting the gap close, and then dash-grabbing to either get the opponent while they're shielding or in hitstun.
One recent training method I've gone with is "bulk input". Namely, exposing the amiibo to as much as possible, and letting it filter through that new information over the course of many matches. Afterwards, I then battle the amiibo to evaluate what it's learned, and assist in 'sculpting' any changes in behavior. Interestingly enough, my amiibo have consistently been more "quick-learning" immediately after a stat or equipment bonus shift, such as Mewtwo picking up D-tilt -> U-tilt -> U-Air only after I'd given it an experimental stat reallocation. Speed stats in particular seem to be the most important for causing an amiibo to re-evaluate their behaviors and become more plastic-minded, as their mobility influences the "what they can do right now" far more than attack or defense, which dictate "how well it works", loosely speaking. First off, amiibo AIs are based upon the normal CPU AIs in Smash. Strictly speaking, an amiibo's data contains weights that modify a normal AI's decision tree. Those terms will need to be explained: Decision Tree: See the Plinko analogy above for a general overview. When a computer of any type is making a decision, it is processing code that looks something like "If (something) is true, do (something), and if that condition is not true, do (something else)". This is called an if-else statement. The game's AI, and indeed basically any AI, is a gigantic chain of these. Weighted Ranges: Think of this as partly-random pattern usage. Decision trees are predictable and very rigid, so any AI-vs-AI situation would be an extremely boring affair as AI are perfectly suited to pattern identification along the same lines as their own way of thinking. The amiibo ability to learn and modify the base game's AI is based in weighted ranges. Let's say an outcome had a base 25% chance of happening among 3 other options, all of which having equal chance of being picked. For a weighted range, this would be expressed as 25:25:25:25. But then say the outcome is a specific move an amiibo can perform among its various attacks. For simplicity's sake, let's say they have 4 valid options, so if they had equal chance of being used, it'd be that same 25:25:25:25 range. Which wouldn't really show any learned strategy, just random behaviors. As an amiibo learns what attacks and strategies (more on this later) work best, that range gets altered. Let's say to 40:30:10:20. That means that when an amiibo has multiple options, it gives some of those options a greater weight in its decision-making process, and over time becomes more and more likely to use the options its training has encouraged. As you can see, the two combine very well to form a Figure Player AI. Determining what option it picks for a situation is not the only way a weighted range could be used, however! Weighted ranges can also influence how an amiibo prioritizes the information available to it. Look at the various factors that play into making a strategic decision: Your damage%, your opponent's %, where you and your opponent are located, what items have spawned and where they are, how the stage may have changed or platforms may have moved, and so on. There's a lot to consider! Which of these are given the most importance changes over time based on what correlate to what outcome. If an amiibo ends up botching a move and getting punished for it, it will have to revise its considerations. So tl;dr, a weighted range is how amiibo prioritize different things or favor options relative to others. Training amiibo largely involves influencing these ranges with results of what worked and what didn't. This is why The Amiibo 15 works. Notice that the guide is very explicit about positioning and focusing upon one thing at a time. The result is that the matches give very strong and clear data for the weighted ranges to be adapted. I won't quote or paraphrase it out of respect for Glenn's distribution. However, for those who have access to The Amiibo 15, you can compare what it instructs you to do with the above information and realize why it works. For those who don't have it, go get it.
Alright, so that's how the amiibo make decisions, but how do they learn? Nobody outside the development staff know exactly, but we can make an abstract guess based on the usual factors: Practicality, simplicity, and fitting into the limited amiibo space. Well, the goal of any match is to win. Working backwards from there, how does anyone win? Hit the opponent(s), don't get hit by the opponent(s). Regardless of other conditions or criteria, that's the core of playing/winning the game. It's very easy to track those interactions! Either damage is dealt, or it isn't. And in the right circumstances, a KO is scored. If a mistake is made, the opponent will be able retaliate with an attack of their own. A poorly-timed shield can be grabbed. A whiffed attack will get swiftly punished, and a KO may result. From these simple principles, learning emerges over time. And with time, more sophisticated experiences and nuances are able to apply. Take for example the way my Mewtwo amiibo learned how to use its Shadow Ball over time, detailed above in the quoted post. Particularly worth noting is the Level 1-50 period for training. This is when amiibo priorities are in a unique state: they seek to imitate you, not defeat you. After level 50, an amiibo adjusts its weighted ranges with error compensation in mind, with the goal of adapting to what defeated them. If they were grabbed and thrown often, they'll use more dodges, rolls, and possibly jabs if they've learned that jabs or other quick attacks are effective at damaging foes who are trying to grab. One can extend these simple rules into application. Namely, if an amiibo learns from battling their opponents, the circumstances of battle and the nature of their opponents will strongly influence what they learn from matches. A simple example would be training on Final Destination vs. an Omega form stage. Normally there's little difference, but then consider characters who can wall jump or wall cling. Boxing Ring's Omega form is a pillar that goes all the way down to the blast line, whereas Final Destination is a roughly trapezoid-shaped floating platform. Boxing Ring Omega will be of much greater benefit in teaching an amiibo with wall cling/jump capabilities to use them! Similarly, stages with blast lines closer or further relative to the stage's main area will encourage horizontal or vertical offenses, such as side vs. up smash attack emphasis. This is part of why Final Destination/Omega stages being featured prominently in training will lead to up-smash spam, as they have closer vertical blast lines than horizontal blast lines. Even more than stage choice, opponent details greatly influence what gets learned. This is obvious. Fighting Jigglypuff will teach your amiibo veyr different things than fighting Ganondorf. A more subtle application of this concept is a technique I call "training wheels". That is, putting on equipment for mirror matches, or giving my current amiibo's opponent specific equipment, that creates a favorable lesson. For example, if I want an amiibo to focus more on dodging and evading, I give them Imperfect Shield for a few days of training. Thus, they become unable to rely on perfect shielding, something amiibo get very good at doing, and are forced to adapt a more evasive playstyle. Coupled with Shield Degenerator, one can teach an amiibo to behave very cowardly or passive as a result. An example "Training Wheels for Passivity" set would be Imperfect Shield, Shield Degenerator, and Auto-Heal. As you can see, that would create very strong pressure on the amiibo to rely on its shield for defense less. Coupled with negative defense stats, and the only option for not getting utterly crushed in battle becomes evasion. Following that set up with a replacement of Shield Degenerator for Nimble Dodger, and the lessons are further reinforced. Another very clear "training wheels" equipment example is First Striker. After a few matches with it, amiibo possessing it become highly aggressive, having learned that as much offensive pressure as possible right off the start of a match is what gives them the greatest benefit. Even without equipment changes, one can set handicaps to matches. In general, a high starting % for an amiibo will make them more cautious, and a high starting % for their opponent will make them more aggressive. In some mirror matches, I occasionally put both myself and my amiibo at 80% to 120% depending on their weight/stats. This is for the sake of teaching them which of their moves are best suited to scoring KOs and which are too slow to be worth the risk. But I mentioned opponents, so will provide example and focus on that. A way to discourage amiibo from using their aerials is to mirror match them while wearing an Air Defender and/or Air Pushover piece of equipment. The result is that they'll see less results from using aerials, and likewise, your own aerial attacks will be substantially less effective. After some matches to adapt to these new conditions and learn from them, the amiibo will use their aerials less. The opposite lesson can be taught with Air Pinata/Air Attacker equipment, for example while teaching Jigglypuff. I've built my Ganondorf to be my ultimate educator, as a result of my findings over time. His moveset is heavily focused on punishes, and with +200 defense and -200 speed, he can ONLY punish, never really press an aggressive strategy, because he's astronomically slow and will never be able to maintain any kind of pressure. Combined with Easy Perfect-Shield, and especially Explosive Perfect Shield, and the result is like a landmine. Poke Ganondorf incorrectly, and you're getting blown away. Keep a safe distance and take advantage of his faults, however, and he can disposed of. One bizarre and amusing result of this is that I only use Explosive Perfect Shield during amiibo spot training to teach the importance of perfect shielding. EPS sends an extremely clear, very strong signal to the amiibo: Use your shield at the last possible frames. Explosive Perfect Shield is better than Easy Perfect Shield for such tutoring, as the latter doubles the frame window, meaning that your amiibo can make mistakes with the equipment bonus absent, while the former simply underscores the benefit of perfect shielding in general. In short, by narrowing what options an amiibo has for securing victory, one directs its learning along that route.
|
|
Cloud
Administrator
Amiibo Dojo Founder
Posts: 380
|
Post by Cloud on Mar 15, 2016 22:59:25 GMT
I really like this post. I agree with everything (except for the up smash spam on omega stages, an amiibo spams up smash if you train it aggressively, not because of the blast zones) and it's very well written out. Great job!
|
|
EPM
Quiet Member
I seek battle, not badges. Pride, not prizes. Mastery, not victory.
Posts: 35
|
Post by EPM on Mar 24, 2016 17:24:09 GMT
3. (Some of) My Own Results
Alright then, it feels only appropriate to go into detail about various findings and training methods of mine regarding the amiibo I own. Not all of my amiibo are worth mentioning. Mii Brawler, for example, is a fighter I virtually never use due to owning Captain Falcon and Little Mac, who both do what Mii Brawler does better. However, it feels incomplete to go into detail about my observations training amiibo, and not provide information regarding...my actual observations while training amiibo.
So I'll be going over some of my more noteworthy amiibo as well as trying to extrapolate underlying patterns or rules!
Mewtwo, Marth, Luigi, Bowser, and Lucario will get varying degrees of focus. My other amiibo have stories and anecdotes, and Ganondorf for example is reliably terrifying, but they're not going to offer anything new here. You all know what Ganondorf can do.
Marth: This amiibo is hard to come by, so I'm going to assume that people have less familiarity with it. With the recent release of Roy's amiibo, I think my experience with Marth (who I nicknamed "Lucina" as a joke) can be particularly helpful.
Marth's setup is Improved Escapability, Critical Hitter, and Improved Dodge. In the past, he utilized Explosive Perfect Shield and Smooth Lander, being the first of my amiibo to be granted the EPS ability when I first obtained equipment with it. Marth is, without question, the most intelligent of my amiibo, and he with Ganondorf and Bowser form my ultimate teaching trio: The Kings. I mentioned in my previous post how I don't use EPS for anything but teaching perfect shielding to my amiibo. This is because it warps other amiibo behavior too strongly. If you read the "training wheels" section, you'll understand why: Most attacks become far more risky if not suicidal, and grabs end up being spammed. Ultimately, however, Marth got so good at Iai Counter usage that he didn't need EPS at all.
I mained Marth in Melee and Brawl, and Marth was one of my first three amiibo along with Bowser and MegaMan. I intended them to be perfect training partners for my other amiibo, especially the then-unreleased Mewtwo amiibo. Mewtwo being my Smash 4 main. So I was able to teach him good usage of Dancing Blade, tippers, and the basic counter move. I gave him Dashing Assault and Crescent Slash as well, finding Shield Breaker to be of less use myself. Crescent Slash is very interesting due to its strange hitbox. Marth discovered that Crescent Slashing at the edge while recovering, if done at the right angle on many stages, will poke an opponent right through the stage. This was one of the first examples I saw of an amiibo discovering something and seizing upon it. It takes quite a while for other amiibo to learn about this, but over time my roster has acclimated to the trick up Marth's sleeve when he's recovering.
Lastly, Marth has seen perhaps the least stat adjustment among my entire roster. I tinker and experiment with others, but Marth has always had satisfactory performance. He's not invincible, but flawless execution of Dancing Blade and scoring KOs with Iai Counter lead me to believe that tinkering is unnecessary for him. For reference, on multiple occasions, I've gone 2v1 with one of my amiibo against Marth and we LOST to him. And I can match him or any of my roster fairly reliably! But sometimes Marth just racks up a ridiculous sum with one or two Dancing Blade chains and then finishes the job with a tipper, and well, I can't even be mad at the squash match because his skill is something for me to be proud of. His AI is reliably competent, and has been picking up combos on its own over time, in part I suspect due to the varying hitboxes of my roster.
Speaking of which...
Bowser: Everyone knows what Bowser does well, and there's no need to repeat it here. Ganondorf isn't even on this list for that exact reason.
Bowser began life with Desperate Immortal, Vampire, and All Around Trade-Off, along with a low-defense high-attack stat spread. I had hoped that I could trigger Desperate Immortal repeatedly, but it turned out that the bonus activates only once per stock. Bowser has always retained Vampire, though his other bonuses have evolved over time. I shun Trade-Off Speedster and All Around Trade-Off in most cases, because the AI is very prone to jumping ludicrously high and just bouncing all over the place. That said, the resulting build was pretty fun to watch and spar with. Low defense meant that Bowser was easily launched, but had the ability to recover from more or less anything below 200%. That meant his weakness of being easily comboed was negated, as even weaker attacks would no longer be able to reliably connect into each other.
Over time, Bowser was buffed, and his merits as a rampaging monster among amiibo were made clear. However, even after tank builds (Critical Hitter, Imp. Esc., high attack/defense, etc.) were introduced, the amiibo did not take to them as much. I suspect that teaching my Bowser an entirely opposite playstyle to what it was raised on for months is particularly difficult.
However, Bowser has one rather silly merit I alluded to: A gigantic hitbox! Bowser has been my go-to amiibo for teaching other amiibo of mine combos. Usually I'll 2v1 with the amiibo to demonstrate the combo to it, and once it starts experimenting, I let the two butt heads for an evening. Bowser's intrinsic properties and frame data make him the best punching bag on the roster, and his own threat means that an amiibo will learn not to just be a berserker on him. But to emphasize my point, Bowser is in my "The Kings" trio of best training partners for this reason. Ganondorf and Marth sharply educate an amiibo about its frame data. Hitboxes, startup and ending lag, What Not To Do, and so on, but of the three, Bowser is the only one who visibly results in actual experimentation and emergent behaviors from other amiibo.
Luigi: Luigi's easily the funniest amiibo I own to watch in action for one simple reason: He confuses the absolute hell out his opponents. I gave him Thistle Jump and Speed Skater (the third bonus shifting over time), and the result is that he moves in such a bizarre way that other amiibo make very silly mistakes when facing him simply because they lack experience against an opponent with such erratic patterns.
Which is a very interesting situation to begin with. Even Marth, Shulk, Ganondorf, and Mewtwo all get caught off-guard when Luigi hits the field, even though they've fought him before. The classic result is them whiffing lots of attacks while Luigi floats like a greased-up butterfly and stings like a pissed-off hornet. That he slides around the ground like an air hockey puck doesn't help, making pinning him down borderline impossible. Luigi's matches take quite a while to finish because he's developed a very patient style as a result of all this.
Naturally, an opponent who moves in a way other amiibo have never seen before is going to confuse them. What strikes me however is that Luigi frequently confuses amiibo who have fought him before. Not all the time, no. But the sight of Marth or Ganondorf flailing around like a level 5 CPU just because Luigi is present is always memorable.
This actually lead to a strategy I've occasionally considered for amiibo tournaments: If none of my opponents have trained for what I throw at them, I have a huge advantage. I've been ineligible for many of the past amiibo tournaments hosted by Dan (or just unwilling to do the team setup), so admittedly I prioritize testing my prize candidates over trying to troll my way to victory. But if amiibo tournaments happened more often, you can expect that inevitably I'll be dispatching something off-the-wall strange in attempt to make other amiibo screw up and make poor decisions.
Lucario: This is less about the amiibo itself and more what I learned from training it, compared to the others. However, I'd like to draw attention first to how the Trade-Off bonuses stack. All Around Trade-Off imposes 60% starting damage per stock. If you have two of it, however, the result is merely 75% per stock rather than 120%. It's like a reverse Law of Diminishing Returns!
What stopped me from doing this? The same thing that repelled me from All Around Trade-Off on Bowser: Pinballing. Not even amiibo with weeks of sparring seem able to handle the ridiculous jumping and sliding this results in. I suspect that the multipliers stack rather than being added together.
What's telling however is that different kinds of Trade-Off bonuses do NOT have their % penalties reduced. I eventually chose to run All-Around and Defender on Lucario, to great effect, and this started him at 90%. However, I eventually went back to 2x All-Around, and paired it with the Overloaded stat spread to counteract the severe movement trouble. He has -200 speed, +200 defense, and +120 Attack, yet due to the All-Around and Aura bonuses, Lucario is freakishly mobile. Coupled with the best Traction stat in the game, and this results in a nasty juggernaut who is good at everything.
Since AWT6 is coming up, I'll share the full build:
Piercing Aura Sphere Long-Distance Force Palm Extreme Speed Attack Double Team
120/200/-200
All Around Trade-Off All Around Trade-Off Desperate Defender
You may be asking why I chose Piercing Aura Sphere. Because he actually charges it. I knew from handling Mewtwo that amiibo dislike charging attacks even when taught to. Piercing Aura Sphere is a concession to this, as it charges so fast that Lucario is willing to make the investment, and more importantly, his damage output on top of the easy spammability (somewhere between Falco's and Fox's Blasters) of the move makes it of far greater utility.
Desperate Defender is the fun one. The effect lasts for 20 seconds per stock. Think about how long that actually is, and couple that with a 200 existing defense stat, and that Lucario starts at 75%. Even my heaviest hitters, short of Little Mac (who is a character I have grown to hate more every time I see him in a match), can't KO Lucario until the 200%+ range, and honestly I'm not sure if that's just because Desperate Defender has worn off by around then. Coupled with a stellar recovery facilitated by the Aura boost, and Lucario is so good I'm still tempted to submit him to AWT6 in place of Mewtwo.
The kicker? Lucario is an amiibo I performed the Trainer Taboo upon: I reset him. I was so displeased with his original self after several months that, deeming him my worst amiibo, I saw nothing to lose in resetting him to level 1. He ended up vastly more competent the second time around in vastly less time. I do not advocate resetting amiibo usually, but this counter-example is quite striking. Data-food for thought.
Mewtwo, being my main and by far the amiibo I've trained the most, will get his own post here another time because I have all kinds of observations. Due to much greater familiarity with the character, I can identify much more.
|
|
Havok
Full Member
Posts: 152
|
Post by Havok on Mar 24, 2016 19:54:19 GMT
Great read. I really enjoy your evaluations.
|
|
Elishima
Undertale Pass Holder
Let's-a-go!
Posts: 918
|
Post by Elishima on Mar 24, 2016 20:04:12 GMT
Amazing, great read as always!
|
|
EPM
Quiet Member
I seek battle, not badges. Pride, not prizes. Mastery, not victory.
Posts: 35
|
Post by EPM on Mar 31, 2016 18:25:24 GMT
4. My Main: Mewtwo (And what I learned from it)
AWT6 will be the first time I get to see my Mewtwo, M.2, in a competitive setting against other trainers. I lack any other Pokemon amiibo besides Lucario, so I'm quite concerned that it lacks sufficient experience against its competition despite my efforts to educate it by selecting them and fighting personally. But hopefully, I'll be given ample data to examine. Hopefully, M.2 will do well enough to make it into multiple rounds. I don't care about winning, I care about the data! I mean, if I win, sure, it's affirmation of my theories and practices, but I value the fight data from observation most. Most of my training has been focused on this amiibo. I main Mewtwo in the game itself, and originally intended for M.2 to be my ideal training partner. Thus, I left it mostly vanilla during 1-50 training (which was 100% mirror matching) except for Imperfect Shield. As I'd trained other amiibo beforehand, I was disgruntled with the AI tendency to just rely on perfect shield timing to score KOs, so intended that equipment to force my amiibo to learn other tactics. Anyway, Mewtwo's training was quite good, though it became clear to me that it is an amiibo with its up and down periods. I experimented the most on M.2, and the sheer number of different equipment and stat combinations provided me with valuable insights into how they affect the AI:
Here's a secret: The biggest influence on your amiibo's AI, as far as its build goes, is the speed stat. Let's look at the Decision Tree / Plinko Board model again. Everything starts with positions. As I mentioned before regarding how the Amiibo 15 works and how amiibo associate their position relative to their opponent (they don't seem to consider their position relative to the stage nearly as much if at all), what actions an amiibo takes depends strongly on their position and what their opponent is doing. I'll even whip up a handy-dandy infographic:
Boxes with green borders indicate choices that the amiibo will make based on training and experience. The choices are there, but which they select is determined by what they've learned. Anyway, since positions matter so strongly in what an amiibo does, and speed stats plus base frame data dictate mobility... The speed stat will greatly influence what an amiibo CAN do, and thus by extension what it WILL do. tl;dr: Speed -> Mobility, Mobility -> Options, Options -> Choices, Choices -> Results. Generally speaking, a fast character is going to end up more aggressive, while a slow one will become defensive simply because it HAS to. So if you want a potential shortcut in making a defensive amiibo, lower its Speed stat. Of course, for characters with high base mobility (such as Mewtwo, who thanks to patches now boasts the 7th best dash speed in the game, above-average walk speed, and air speed in the top 4) will remain aggressive. Frustratingly, spot training an amiibo tends to make it more aggressive even when I'm specifically trying to teach it defensive preferences. I, and no other human, will ever be able to react at a consistently equivalent speed to an AI. I'm competent and can beat my amiibo about as much as they do me, but my reflexes will never match them, and so they pick up the bad habit of 'assuming' that behaviors capitalizing on that will tend to work out. Which, against other amiibo, will fail miserably due to being matched.
Back to Mewtwo. I already detailed Shadow Ball in post 2. Over time, I amassed enough equipment to match my amiibo acceptably and craft training partners for it, so it ceased being vanilla. During 1-50 training, I taught it to prioritize the following moves: Down Tilt F-Air Grab Shadow Ball Up Smash Mewtwo's moveset has more gems than that, but I took a wedge approach to my priorities, as I tend to with almost any system. In other words, I'll pick a top priority, but not specialize in it to the fullest extent so I can also tick up some secondary or even tertiary priorities. Sure, not as powerful as a specialist, and not as versatile as a balanced build, but I enjoy having options and can accept weaknesses if I'm able to adapt to them. M.2's build? A secret until AWT6! TheEleh knows a morsel of secret tech we both share, but the submitted build is changed. For example, I didn't include Improved Escapability. "Why, everyone here knows Cloud's a huge fan of that ability, and it's proven to be very effective, that's crazy EPM!" Well, Mewtwo has some of the best throws in the game. If you're in range to grab Mewtwo, Mewtwo is in range to grab you. And Mewtwo's grabs are a lot nastier than most of the cast's. Charizard's throws compete with Mewtwo's for some of the strongest in the game, and Lucario benefits from the fastest pummel in addition to Aura boosts. I'm taking a gamble with some of my choices, but the training was so promising that I'm willing to bet dangerously. I've always been reckless...
Anyway: I'm glad someone else can attest that Mewtwo can be trained out of Teleport-SDing. Anecdotes of Mewtwo only doing so off one side of the stage are particularly interesting. One thing I noticed when mine did this (and it still happens rarely) was it would do so away from me, never in close proximity. What I think is the root cause for "Telecide" is that Mewtwo's AI is trying to experiment, as amiibo always do. Even heavily-trained amiibo will deviate in their habits for the sake of trying an alternative out to see how it works. The "medium" range, about half the distance of Final Destination, with Mewtwo in the air, seems to be the trigger condition for Teleciding. This is unsurprising, as this is the most uncertain range for almost any fighter. Think about it: The sheer number of motion options present a higher chance of making a poor choice, the situation is dangerous as it's very easy for an opponent to either juggle them or punish their landing (such as with a FC Shadow Ball in mirror matches), and the only option for most fighters is down-air or a relevant special move to try and ward off a foe. Teleport is a versatile "get out of there" option, with the bonus of being intangible during its movement. The end result is that Mewtwo prioritizes escaping a dangerous situation (especially given its low weight, generous hitbox, and floaty jumps), Teleports away from its opponent, and proceeds to self-destruct. If I had to guess, the AI is lacking some kind of flag on Teleport that says "check to make sure there's a surface to land on below the teleport destination", and has to learn that rule implicitly as a "don't use Teleport in this and that direction from these locations relative to the stage". Experience will eventually groom the issue out to a fluke rather than a habit. I advise the Umbra Clock Tower stage to educate a Mewtwo amiibo about its teleport usage, as the secondary platforms shifting, while the primary stage remains fixed, is ideal. Additionally teach it to teleport straight down to the ground or only at slight angles, rather than at full diagonal angles or sideways, when on the stage. No more than 45 degrees from straight down. This allows Mewtwo to escape the threat of being juggled, which it is quite right to consider, without chucking it off the stage for a stupid SD. Teleport has significantly less landing lag if it ends on the ground, rather than Mewtwo going into helpless state in the air and THEN landing. Combined with air dodge training and aerial Confusion usage as a deterrent, and Mewtwo can learn significantly better alternatives to what it's trying to do when Telecides happen. I feel like that info should be in a separate thread, but I'll cross-post it in my Research Compilation in the future. Over time, and I mean several months, M.2 has learned some pretty nasty combos. To my surprise, it can and will execute a D-tilt > U-tilt > F-air string if it gets a chance to, and one very surprising thing that made me suspicious was how it learned to drag opponents off the stage or up off the ground with N-Air!
Which is something I could never pull off, let alone while mirror-matching it. It literally never saw anything like that, and then suddenly started doing that. Usually it'll whiff the upward-pull maneuver, especially on me during mirrors because Mewtwo has the best air dodge in the game. But any time it yanks Ganondorf off the ledge, then teleports up to safety while Ganondorf flails around trying to feebly recover, is hilarious. It's not consistently effective, but it is consistently attempted now and then if it has the opportunity. I knew Mewtwo could do that from Smashboards combo threads, so realized what it was trying, but I know for a concrete fact Mewtwo was never exposed to the technique in any way. It even tries to follow up the pull-upwards with a F-air. If at all possible, I'd love to teach it to do D-air instead with that, but hoo boy that'll be hard. Additionally, a combo I never knew (and cannot pull off due to the frame timing being ridiculously strict) is Confusion into F-air. This was something Mewtwo started doing even before it hit level 50, and I actually emailed Cloud about that back when he was working on the Mewtwo guide. So with those factors, I've concluded that amiibo have a hidden "experience" counter of some sort, possibly a composite of several factors. Number of battles fought, time in-battle, amount of damage given/taken, whatever isn't the number of reads/writes to the NFC as covered in post 1. As this increments, it'll automatically pick up new tricks. One could compare it to an RPG character learning new abilities as it levels up. This also falls in line with the "simplicity" criteria of amiibo structure. There's no question amiibo learn, but in addition to that, it seems Nintendo hard-coded at least some elements of progress over time, likely to provide players with the sense they were improving their amiibo regardless of whether they were a 9 year-old or a seasoned player. After all, Game Design as a whole is all about user experience, and I hate to be the "Santa isn't real" guy, but a lot of game content is designed to make you feel competent without actually requiring you to be. For example, "non-puzzles", or any sequence where the actions to take are clear, but the actions themselves require an obvious order without actually being instructed. Since the player reads nonverbal cues and completes the task, there's a happy little dose of neurotransmitter glee at "yay I solved a puzzle" without their actually having been a puzzle. Granted, this is subjective. A less-murky example is difficulty scaling. Even in games without multiple difficulty settings, or games that only change them at the start or in a menu, it's pretty common industry practice to sneakily make things easier (or harder) on the player to prevent too much frustration or boredom. The trick is that this has to be done without them noticing, or else they'll feel bad about it. For example, random factors start to take on a skew more in the player's favor, and you can bet that weighted values play a part in that. Time spent on an objective, number of retries, etc. can and do play a role in "luck". Or say there's a really nasty boss fight the player keeps dying on, in which case they might use their Really Annoying Hard To Dodge Attack less often while still keeping the player engaged with their Reasonable Reaction Time Attack, possibly slowing down a little or moving around more to create the illusion of the same pressure and activity. Health drops might come more often, enemies might get worse aim, so on and so forth. Once again, I'm not saying amiibo are like that in entirety, but I do not doubt that specific sequences of actions or advanced tricks are built into every amiibo, and are unlocked over time somehow. After all, if it was just about the frame data, veteran amiibo would be seizing on any window of combo opportunity and things like THIS would eventually happen:
|
|
Z
Full Member
Now I am become $wag, the moistener of your mothers' granny panties
Posts: 223
|
Post by Z on Apr 3, 2016 18:01:06 GMT
Whelp, we'll get to see whose training methods prevailed tonight. I'm actually scared since you probably play Mewtwo a lot better'n I do and have a more refined training method. I just pitted Gyiyg up against my other pokemon amiibos between mirror matches and team-up fights. He was very hit-and-miss, but he did manage to 0% 2-stock Lucario once. Good luck to you, I'm expecting quite the riveting mirror-match this evening. Edit: GG, bro. M.2 fought bravely.
|
|