Is the graph of the top a representation of the neural net? That is, are 11 neurons in 2 layers enough to operate the game?
I too am perplexed as to what the neurons represent, the blog describes the RNN as 19 inputs connected to 7 outputs.
Examining slimevolley_pro.js (Agent.prototype.drawstate) shows only the agent states are depicted onscreen x,y,vx,vy,bx,by,bvx,bvy in the top row and actions forward, jump and backward below.
The code is well documented and builds on convnet.js by Karpathy.
The network is trained by genetic algorithm and self play - so this is a neural net trained with reinforcement learning.
The author's method seems unique and effective.
The blog post comparison of the resultant 'genomes' of network weights seem to show that the space has been searched exhaustively.
The game is played with trained networks but the code contains a training flag - it will be interesting to watch them train with self play.
http://blog.otoro.net/2015/01/27/neuroevolution-algorithms/
He outlines the evolution of his thinking and slime volleyball in this post which cites John Gomez's thesis as the inception of CNE.
Certainly parallel ideas to Schmidhuber but the implementation details are somewhat different in the U of Texas Neuro-Evolution models.
http://people.idsia.ch/~juergen/evolino.html
Yet Schmidhuber's nets are much more complex and certainly different.
I still think Otoro's very simple RNN feedback nets are unique - especially when coupled with training by self-play.
Does Schmidhuber have any game playing agents ?
http://www.cs.toronto.edu/~graves/handwriting.html
Using a recurrent architecture.
To be honest I haven't played with NNs, but it puzzles me as to why the non-recurrent approach is so prevalent for complex tasks. I mean, it's the basic combinatorial circuit vs sequential circuits, which we all know are much more suited for complex or large outputs. Where's everything we learned from synchronous logic synthesis?
Indeed this is exactly it, evolution is a global method, learning is local.
I am reading John Gomez's thesis where he compares and combines learning and evolution
http://www.cs.utexas.edu/users/nn/downloads/papers/gomez.phd...
Otoro's post on the evolution of his slime volleyball thinking is well worth a read.
The connection comes up in david mckay's Information Theory book too, a reading I definitively recommend, although I haven't been through it properly myself.
Having learning couched in Information Theory terms brings it all right back to Claude Shannon's early work on Reinforcement Learning Chess programs and Alan Turing's ideas about evolving efficient machine code by bitmask genetic recombination.
Backpropagation takes longer the deeper the net - Recurrent Neural Nets are deep in time so Back Propagation can become intractable or unstable.
Otoro's Slimeball demo evolves a Recurrent Net rather than training it - this appears to be a very efficient method, less likely to get stuck in local minima.
The slimes evolve through self-play which is a trial and error method and reinforcement methods seem to do better on control tasks than passive learning.
I wish I stayed at U of T for a few more years...