Disclaimer: I was tempted to put this in Game Design, but it heavily references AI so feel free to move it to wherever it needs to be if this is the wrong place.

I had this idea that a game could use machine learning to support it's AI and make itself a really challenging opponent (nothing new there), but also to tailor its style of playing based on feedback given by human players.

Using an RTS as a classic example, lets say you prefer to play defensively. You would probably get more enjoyment out of games where the opponent was offensive so as to challenge your play style. At the end of each match, you give a quick bit of feedback in the form of a score ('5/10 gold stars' for example) that pertains to your AI opponent's style of play. The AI then uses this to evaluate itself, cross referencing its score against previous scores in order to determine the optimum 'preferable' play style.

Then I got onto to thinking about two issues with the idea:

1) The human player might not be great at distinguishing feedback about their opponents play style from feedback about their game experience in general.
2) In a multiplayer context, players could spam/abuse/troll the system by leaving random/erroneous feedback.

Could you get round this by evaluating the player without them knowing it, i.e. could some other data recorded from the way a player acts in a game be used to approximate their enjoyment of a particular opponents play style without being too abstract? For example 'length of time played somehow referenced against length of time directly engaged with AI opponent' etc...

Do any existing games work like this? I just came up with it when I saw a stat that call of duty has been played for a collective 25 billion hours or something - which made me that would be the perfect bank of experience to teach a deep learning computer how players interact with a game.

Just a bit of abstract thinking, that's all.

Almost every mobile game on the market is collecting session data and using that to inform decisions about how to tweak the game experience, including deliberately giving players differing experiences and measuring which work the best in terms of session length, player retention, monetisation, etc. Is any of it using machine learning? Probably not, or at least very little, because it's not necessary - this is usually just a simple regression or statistics problem.

I think it's important not to think that because we're talking about AI in the game sense, that tweaking it requires AI in the 'machine learning' sense. Game AI has relatively few inputs and outputs and as such you don't need complex machine learning capabilities to be able to see the relationship between the two sets of data.