I was just poking around for interesting papers on Google Scholar and I saw some pretty good results in this paper. The best reinforcement learning neural network got a 99.95%+ success rate in 3/5 scenarios, 99.75%+ in 4/5 scenarios, and 98%+ in all 5 scenarios.
Apparently the other cars are controlled using an algorithm developed in 2000, so they might be pretty dumb. Performance might be better with smarter agents in the simulator.
Here’s the second experiment from the paper. In these next scenarios, the “non-ego cars are not allowed to decelerate” and “do not brake for the ego-car”. So, the ego car is essentially playing Frogger. The ego car also can’t see many of the cars that are coming. Kind of a crazy contrived situation.
Please share more papers, talks, etc. about reinforcement learning.