Is there a suitable traing enviroment for “Director Class” game ai


#1

Hello, everyone
What i mean by director class is a non npc ai that has some abilty to control in game events. or is able to in a sence ie, a good example would be left 4 deads main directer ai. Controlling item spawns, mob spawns, etc based on the players progress/stress lvls, overall budgets of mobs etc and the general mood the game is taking and keeping it fresh with highs and lows. I was wondering if an environment exists or else could be created to train this type of ai
Thank you


#2

There cannot be a fully automated environment for this with current state of AI. The reward signal would be player emotional state and engagement, which cannot be reliably simulated.

There are level design AIs, which are considered a form of computational creativity (maybe also procedural generation), and these could in theory be filtered through a learning agent, in order to meet design constraints - e.g. maximum jump distance, non-trivial yet not impossible placement of objectives. Or in reverse, having an agent somehow assess the playability versus a particular director by learning to play against them.

The last part gives one possible avenue of research. You could run a director agent against player agents, with the objective that the players somehow score how much “fun” they are having whilst playing, and at what player competence level. However, at the moment that would need to be a very manually tuned setup, since the AI players cannot provide any feedback on subjective experience. You would have to use proxies (e.g. variety of actions/behaviour scored according to some cost or utility function) and careful routine benchmarking versus real human players.