hardmaruReynolds steering agents trained to avoid moving walls. Brain is a recurrent neural net evolved via genetic algorithms where the few surviving agents pass on their code to the next generation
- jianyshenNice! Looks like a variant of the convnetjs rl demo. I noticed you're using recurrent nets for most of your work, is there a way to quantify the benefit of recurrent nets in these simulation settings over a deeper feed forward net w some sort of pruning?
- hardmaru@jianyshen thx- this type of really simple RNN is equivalent to a feed forward nnet with infinite depth, so in a sense it is by definition deeper than any feed forward net, but also requires much less memory and connections. I found RNNs good for temporal difference problems where patterns occur in difference of time, similar to control systems with feedback states, where as ff nets may be better for classification sort of problems.
- jianyshenOh ok, it would be far less parameters. Have you tried to interpret the trained network to see what what features and corresponding strategies your agents actually learned? Is it fairly straightforward in these cases or not necessarily so? Btw the combination of using js and processing is awesome for the visualizations
- hugehugleIt let me think of catching chickens when I was child:)
- hardmaru@ctrlcollective @wulongt thanks!
Log in to like or comment.