You are always training an AI

I’ve been meaning to write on this ever since I first trained my Sheik, so this isn’t news to many people in the amiibo training scene. While I’m pretty sure this was also the case in Brawl and Smash 4, Ultimate’s CPUs seem to exhibit the behavior I’m about to describe to a much more significant degree.

A while back, I wrote a post about a Sheik I had trained, titled “Why Sheik is a perfect example of the Utility formula at work”.  In this post, I described the Sheik in question: I had taught her to use forward tilt and forward air repeatedly in order to juggle the heavyweight opponent off-stage. The training consisted mostly of using just that move for a very long time, and resulted in the Sheik largely using that option as well. The training matches often ended with both of us spamming those two moves, and nothing else.

Prior to this, I had only used Sheik for four matches, according to the game’s built-in records. Other than those four matches and the amiibo training, I had never played Sheik on my Switch console.

After completing my Sheik training, I was curious about how effective the forward tilt, forward air method would work on a normal CPU. I was still deciding whether to enter her in tournaments, and wanted to poke around and see what I could come up with. I set up a Sheik ditto against a level 9 CPU, and used the method I had taught my Sheik: forward tilt, forward air combos in some varying order. Nothing special. But to my surprise, before I could even start the combo, the CPU was using it against me!

61qWAL2v6ZL._SX425_

This continued on for a few matches before I decided to stop. I hadn’t observed this behavior from the level 9 Sheik AI before, but it wasn’t something I had known to watch for. I found some footage of the level 9 CPUs on various Youtube channels… but no behavior like what I had just seen. (A month later, I observed the same phenomena when training my Shulk amiibo.)

In addition, I noticed that training an amiibo of a character I don’t often use causes that level 9 AI to fight more like that amiibo. I spend a lot of time fighting against CPUs (I’m good enough at Smash to beat my friends but not good enough to compete) and it really seems that the specific behaviors I teach to my amiibo have been adopted by the AI. This isn’t exclusive to training amiibo, either, as it appears to me that characters for whom I’ve never trained an amiibo also end up with the AI mimicking my playstyle. At its simplest, it seems that when a human is playing a character, the game is taking note of their play and adapts that to the 9’s AI.

I theorize that the specific phenomena I observed with Sheik and Shulk occurred because the game has no other inputs from which to derive. I didn’t play Sheik at the time, and I’ve not played Shulk. The game is  programmed to have its own base AI for Sheik, but it also is supposed to adapt to what that character eventually uses and  overwrite that base AI slowly. Similar to an amiibo, it will learn over time, but unlike an amiibo it has a different base AI (free of the quirks of its respective amiibo) and is learning from you every time you use that character.

My sense is that in Ultimate, amiibo’s base AI and that character’s CPU AI are tied together in some way, at least until the amiibo is trained and has adopted its own playstyle. It also looks like the AI in Ultimate, much like amiibo AI, is not 100% flexible and it has some functions hardcoded in: recoveries, using specific moves, etc. However, amiibo still has its own similar but distinct base AI and quirks that the CPU AI doesn’t have.

In conclusion, it seems that over time the CPU AI learns from human inputs and adopts the playstyle to some degree. This is dependent on your specific save file, and it can be demonstrated by teaching an amiibo a very specific behavior. (I recommend amiibo because they stand still, especially if you use the FTFW method, which I will be writing about soon). It’s unknown how this affects amiibo tournaments, or even if it affects amiibo at all. It’s such a specific behavior that there’s really no way you could claim the placebo effect, either. At this point in time it really seems that CPU AI learns from human inputs.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s