How do I create an AI controller for Pacman?

Question

How do I create an AI controller, which can play pacman - by taking in pixel values (or some other data by represents the state) which perhaps runs on a separate thread, which can control the game?

It takes the pixel values (or likewise) to learn, and play the game, just like a human does. Normally, AIs like enemies - are part of the game code, which learn and play by using the internal state of the game, just like the “enemy” AI.

How can I create this abstraction between the game and the agent?

The most obvious example would be : (I don’t know how viable this is) - The agent takes in pixel values of the game - learns - and sends “keyboard” inputs to that thread just like a human would, to play the game.

Can someone point me in the right direction - as to how to do something like this?

Answer 1

About the environments
For the controller part of your question, I would advice looking at openAI gym. https://www.gymlibrary.ml/content/environment_creation/ #how to make your own gym enviroment
Those gym enviroments work kinda like this

env = gym.make("ALE/SpaceInvaders-v5")
observation = env.reset()

observation, reward, done =env.step(action) 

Where your observation can be pixels,ram dumps, etc. The actions can be internally mapped to key presses. (but the internals don't really matter that much) Gym environments are just an easy way to then have your agent take in the observations and map them to actions. So your game can be stepped through so you don't have to worry about a complicated way to integrate the AI and keeping it synced.
There are already a bunch of artari games included (also ms pacman) https://www.gymlibrary.ml/environments/atari/

About the AI

Since you want to learn directly from the pixels or other more complex data types I advice you to read this:

Deep Reinforcement Learning, a textbook (arXiv:2201.02135) https://arxiv.org/abs/2201.02135

If you just want some simple tutorial just read "Hands-On Reinforcement Learning with Python", then you can try to implement something like DQN , with some cnn architecture. (similar to how they did in arXiv:1312.5602 "Playing Atari with Deep Reinforcement Learning"). Since DQN is model free, off policy and relatively easy to implement. (cool thing about off policy is that your agent doesnt have to interact with the environment as it doesnt need to sample "experience" using only it's current policy (way of choosing actions given a state) so you can even play the game itself and collect information (a,s,s',r) while playing and train your agent on that (to nudge it in the right direction if it gets stuck))

if you're not a fan of gradient descent based methods, you can also use methods like ES (evolutionary strategies). As that way you can directly optimize for a reward. (it's a genetic algorithm that scales well in both compute and dimensionality) (arXiv:1703.03864 Evolution Strategies as a Scalable Alternative to Reinforcement Learning)