Bold title ..... - 3 views
-
johannessimon81 on 02 Sep 16I got a fever. And the only prescription is more cat faces! ...../\_/\ ...(=^_^) ..\\(___) The article sounds quite interesting, though. I think the idea of a "fake" agent that tries to trick the classifier while both co-evolve is nice as it allows the classifier to first cope with the lower order complexity of the problem. As the fake agent mimics the real agent better and better the classifier has time to add complexity to itself instead of trying to do it all at once. It would be interesting if this is later reflected in the neural nets structure, i.e. having core regions that deal with lower order approximation / classification and peripheral regions (added at a later stage) that deal with nuances as they become apparent. Also this approach will develop not just a classifier for agent behavior but at the same time a model of the same. The later may be useful in itself and might in same cases be the actual goal of the "researcher". I suspect, however, that the problem of producing / evolving the "fake agent" model might in most case be at least as hard as producing a working classifier...
-
johannessimon81 on 05 Sep 16This paper from 2014 seems discribe something pretty similar (except for not using physical robots, etc...): https://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf
-
Paul N on 13 Sep 16Yes, this IS basically adversarial learning. Except the generator part instead of being a neural net is some kind of swarm parametrization. I just love how they rebranded it, though. :))