5 comments

[ 2.8 ms ] story [ 15.6 ms ] thread
This is an excellent article with a great point.

However, I don't think it's a bad next step to train AI's to play game.

Even though games lack several of the difficulties of real world tasks, playing them well is just beyond our current capabilities. It is good to successively solve hard problems, rather than wait for the perfect unified answer to everything thing.

Who suggested waiting, in any shape or form?
We just have to avoid getting stuck at a local maxima of AIs that are good at playing games :)
This video of a Super Mario playing AI was pretty neat:

https://www.youtube.com/watch?v=qv6UVOQ0F44

It may, or may not be on a useful path to solving harder AI problems. I don't know. But it's a good way of demonstrating value (and publishing papers).

The article doesn't really focus on algorithms, but rather reward functions. In this case, the evolutionary algorithm has a reward function that tells it which mutations are most useful and therefore should be used to generate new mutations.

This reward function is essentially the same no matter if you do normal reinforcement learning or evolutionary.