It would be interesting to see if you could improve on the results supposedly disproving having an 'objective'. I haven't sat through the full 40 minutes of the

It would be interesting to see if you could improve on the results supposedly disproving having an 'objective'. I haven't sat through the full 40 minutes of the video and I don't see any writeups on the Picbreeder website (which appears to have not had any news updates since 2011), so I am guessing that what he did was using a CNN image classifier and do greedy maximization over the various options on each step; it will surprise no one in RL that this does not work as it is not a convex problem. But a RL CNN trained to instead predict human choices (imitation learning) or do RL learning with rewards just from the final step (perhaps Christiano's preference learning setup?) might do much better. After all, all the comments about the folly of greed apply just as well to choosing moves in Go, but you don't need to abandon 'the myth of winning' to win Go. It's simply about having an excellent heuristic for promising directions and keeping options open.