This thread is for anyone working on personal projects to share their progress, and hold themselves somewhat accountable to a group of peers.
Post your project, your progress from last week, and what you hope to accomplish this week.
If you want to be pinged with a reminder asking about your project, let me know, and I'll harass you each week until you cancel the service.
Jump in the discussion.
No email address required.
Notes -
Recently somebody here, I think @RandomRanger, talked about using AI-as-in-LLM to build AI-as-in-game-opponent.
There is an old simple turn-based game that I've been playing for years, but I have to treat it more as a solitaire, because the computer enemies are frustratingly weak even on the highest difficulty, and it's very uncommon to struggle with a map. Except there is one infuriating quirk, where the enemies gang up on the weaker of the leading sides instead of the strongest, despite the game mechanics being very snowbally, and so the hyenas never even prosper themselves, only make a player's stumble almost never recoverable. (The recoveries are the most memorable matches, though.)
Claude Fable graciously one-shotted a clone of the game for me (and I wasted an embarrassing amount of time playing all the sides by myself) with a prompt that stipulated it should be decoupled enough to support both a human UI and a connection for artificial opponents.
How would I go about producing strong, motivated to personally win opponents? Ideally it would be something like AlphaGo self-playing training, but I'm not sure if that would be feasible on a consumer desktop with a middling GPU.
Well you can use evolutionary weights (CPU-based), just running the game for many generations and having it become more adept via self-play, evolving a series of parameters that affect its decisionmaking. But this then results in various causal issues, like the AI not understanding that X needs to be made before Y. Or you can try for search, where the AI scans all possible (ideally pruned) paths and aims for the best. Or a combination of the two. Some parts of the game may be more amenable to search than others, search is by definition the strongest approach but gamestate branching makes it impractically expensive unless cleverly managed.
AlphaGo does more complex things with neural networks that might be hard for a consumer GPU. That approach wasn't suitable for my game due to a certain level of complexity (and that I'm adding more content and rules over time) but maybe it would be for yours since the rules are unchanging.
Fable would know more than me. Ask Fable for a detailed plan.
And also work out a good statistical base to do A/B tests so you can see if the AI is improving much. I'm watching these tests happen right now on my other monitor. It's an interesting process but once you start you begin to appreciate why strategy games have such poor AI, it's like dipping into an ocean of limitless depth.
More options
Context Copy link
More options
Context Copy link