Revenge of the Nerds! Humans Beat Bots at Dota 2 International | Artificial intelligence

I'm actually happy that the AI lost.

I feel that the SoA RL methods still have some limitations that are hard to identify because you can get over them with “enough computational power”.

I think that's exciting that there's a challenge that might require the current RL methods to evolve.

I've heard (not sure if this is true) that the OpenAI team communicates by sharing some neurons in a recurrent network. That's a cool solution, and it seems to work well enough, but this seems like an open problem with a lot of possible solutions.

I was also intrigued by the amount of times the AI went to Roshan to check if it had spawned. Learning respawn times doesn't sound too hard, but the AI seems to be losing track of time. Roshan takes so long to respawn that it might not be feasible to solve this simply by increasing the recurrent net memory, so maybe new methods could help here.

Also, this problem is simple for because of the “clock” shown in the spawn point. How hard would it be for an AI to learn how a clock works? Could transfer learning help here?

I'm really excited to see how OpenAI will tackle those problems. I just hope that they aren't solved with just more training time.

You might also like

Comments are closed.

This website uses cookies to improve your experience. We'll assume you're ok with this, but you can opt-out if you wish. AcceptRead More