Batu Aytemiz

It is extremely rewarding to approach games with an analytic eye, trying to push the field forward in both technical and theoretical ways. I am working on reinforcement learning, which tries to mimic how people and animals learn. Roughly speaking, I am trying to create artificial intelligence agents that learn to accomplish a set of goals through discovering what works by trial and error, instead of explicitly coded instructions.

I have two possible ways of going forward. One is to create a game playing agent that takes in pixel data, similar to what we see as humans on a screen, and then learns to play the game. This is not new, as agents have been released that can play games pretty well. However, the reward function, or what it means to be good at the game, was defined by humans. I am looking into ways for a machine to understand what it means to be good, and then using those assumptions to try to maximize its gameplay.

Another way forward is to see whether we can embed emotion to the playstyle of our reinforcement agents. Is it possible to play chess aggressively, defensively? What does that look like in a computational setting? How can we tell a machine to learn how to play with a certain emotion?