Short video for context: https://twitter.com/P1xBoi/status/1610045154760458245 I have this platformer game where you go about procedural levels, trying to find the end. I've been recently wondering if ML could be used to place enemies intelligently around the level, so they create a challenge for the player. What my question is that how can the "enemy director agent" observe the level? I have a node thing with dictionaries and graphs, containing information about the node, rotation position, type etc. But observing this all would create for a very large observation space? Or is there something I'm missing? Developing a reward scheme for this kind of thing seems very tricky. For now I've been using graphs to calculate the "most optimal" route for the player so that all areas serve some function, placing trees, obstacles - but this seems like a very dumb approach in that way that the levels are only as smart as me. My dream would be that I would have some kind of adversial scheme of three agents, the player itself, a devil that tries to make the level hard, and an "angel" that would help the player - but this is probably too challenging and hardware costly to implement. If you have any ideas please comment. I'm a beginner in ML so I might be overestimating the potential of ML altogether, if so, please point out!