manny405 / sapai

Super auto pets engine built with reinforment learning training in mind
MIT License
67 stars 21 forks source link

Project activity #77

Closed andreped closed 2 years ago

andreped commented 2 years ago

Is the project dead? If not, it would be fun to contribute or at least test it out myself.

I had the idea of developing a bot using reinforcement learning after being trash at the game for too long.

One thing I was thinking would be a challenge would be whether I would get access to real games, lots of them that is. Such that I could develop an AI that could beat real people, or at least improve itself in specific situations where the player made a mistake.

Do you know if there is a public database where games are being recorded and stored in an easily machine-readable format? That would be a great foundation for developing the bot.

manny405 commented 2 years ago

Perhaps not dead, just in a deep hibernation... The pets and their behaviors have been fairly stable for a while and it's a better time for me to hop back to the project, so I've been considering another push to get back up and running.

I do not know of datasets of real games, although I also have never looked. It's not a particular interest of mine unfortunately. I'd rather start the bot from scratch compared to using previous human game data because I have a lot of compute available to me.

andreped commented 2 years ago

so I've been considering another push to get back up and running.

Great! I am interested in contributing, as I'm having fun creating AIs for SAP. Are you contributing to sapai-gym as well, or is that mostly @alexdriedger ?

alexdriedger commented 2 years ago

@andreped I was the main author of sapai-gym, and it's functioning as a gym wrapper over sapai. If you're looking at doing RL, I would highly recommend trying it out if haven't already.

I got it working to a point where I could train an AI with RL for the base set of pets, the AI was just wasn't very smart (couldn't beat my baseline agents). I'm not an expert on ML, so I kind of left it at that point, but I'm sure some tuning and improvements could get it further

I am not actively working on it anymore, so if you would like to make changes, forking it is probably the best option. I'm happy to answer any questions about the library (making an issue is probably the easiest route there)

andreped commented 2 years ago

If you're looking at doing RL, I would highly recommend trying it out if haven't already.

Already started using sapai-gym, works great! Haven't had the time to properly assess its performance, but reward-wise it seems to be improving.

I am not actively working on it anymore, so if you would like to make changes, forking it is probably the best option.

I have already created a fork. Will make a PR soon, and create some issues if I find any breaking bugs. Will you be able to address these PRs/Issues if I make them?

alexdriedger commented 2 years ago

Fantastic to hear that the library is working well for you! I can respond to any of the issues you make and merge the PRs on sapai-gym (assuming they look good 😅)

andreped commented 2 years ago

assuming they look good 😅

I will do my best :P

manny405 commented 2 years ago

I'll close this issue. We can continue more in new issues or discussion threads.

andreped commented 2 years ago

I'll close this issue. We can continue more in new issues or discussion threads.

Good idea. I made this issue as there was no Discussions tab.