Machine learning in games

Jurgen

This is something I'm very interested in.
As something for the future I would like to implement Machine learning into my games.
I currently work with Game Maker Studio.

Does anyone have experience with implementing Machine learning into their games?

Any help or info will be highly appreciated:)

Here is a video of Machine learning in action:

francoisvn

Machine learning is such a broad thing so the answer is pretty much: it depends on what you want to do. You could want to use it for an AI player, or for procedural generation as part of a generate and test meta algorithm, or something else entirely. I have some experience with machine learning, but it's not the same as MarI/O, so if that's what you wanna do it'll probably be of little use to you. Out of interest's sake, I used reinforcement learning for a Go player.

That said, machine learning is also a very deep rabbit role you can fall into. If you're wanting to do it as a learning exercise I can see the value in that, but there are likely much easier ways to make a game better. For example, if you wanna make an AI player, machine learning is unlikely to be the best approach except for a research topic, the smoke 'n mirrors approaches usually perform better for a fraction of the implementation costs.

Jurgen

Cool:) Thanx for the feedback @francoisvn. You've also answered another question I've had.
I've been thinking of a way to improve fighting game AI where at the moment I am using a Finite state machine.
An idea I've had was to get the AI to learn to execute better/ more efficient states as the fight progresses.

I could do this without Machine learning I reckon. If I take note of player input over time, I could adapt the AI's behaviour.
For example: If the player falls into true for 'Player-agressive-behaviour' then the appropriate behaviour for the AI 'Ai-counter-behaviour' will be executed.

francoisvn

From my limited knowledge of fighting games I imagine the AI is a whole field to itself. Maybe there are some talks and/or articles on that? If not, it might be a good idea to try model good human players' techniques. If you have a few slightly different models or parameters for players that your AI can perform, you can do some opponent modelling by doing constant little predictions of what they would do based on the diff models and then assuming the model(s) that is the most accurate is a good approximation. Once you have a good opponent model you can have a pre-constructed matrix of what model works well against what, and try adopt a model that works well against theirs. This is mostly just my idle thoughts and this is all easier said than done, but maybe it's an avenue down the right path? It's likely to be considerably easier than many machine learning techniques, cause you'll mostly be dealing with heuristics and building custom behaviour models instead of building deep neural networks.

ProjectX

Hey @Jurgen check this out: http://dl.ifip.org/db/conf/iwec/icec2007/ChoPY07.pdf
I can't comment on the usefulness of this article, as I cannot read it right now.