Hi,

First of all excellent article as always very pleasant to read.

Clearly this is a burning issue, I’m agree with you about the fact it’s may be not the good strategy to take to solve RL problems BUT this strategy has the advantage of call into question our conservatism in RL.

What I mean about that is the fact that when you read about the history of AI you see that during a long time there was only one way to do AI it was hard coded knowledge. But some people dare to say “what if we use another strategy” a lot of these new strategies were really really bad but they had the benefit to destroy the conservatism in the industry and lead to the resurgence of neural networks.

Sure I’m totally agree with you about the fact that imitation learning is a simple answer to a complex problem. Maybe this strategy will lead to a new set of awesome strategies. But, for now I prefer to look at OpenAI retro contest results to see real improvements in RL.

However, I really hope that because it’s OpenAI and DeepMind that published papers about this it will not lead most researchers because of the hype on working only on imitation learning strategy.