From: joshtrask@gmail.com   
      
   Upon doing a lot of reading, the clearly obvious has become even more   
   obviously clear: there are many ways to do neural nets, and there are   
   many ways to do reinforcement learning. What's less obvious, however,   
   is how to combine the two. Neural nets are obviously going to be   
   better than something that just manages (state,action)->(reward)   
   lookups at recognizing correlations between state and reward, so it   
   seems natural that for complex RL tasks, the software should be   
   powered by a neural net. Aside from the one used in Grand's   
   "Creatures", however, I've been unsuccessful in finding combinations   
   of the two - techniques for adjusting the weights in the net in   
   response to the reinforcement signal. Though Grand's seems good, I'm   
   looking to enumerate all of my options. One that occured to me would   
   be to use a standard algorithm (probably recurrent) to predict the   
   reward for a (state,action) pair and then using something like   
   Q-Learning to evaluate what to do with the predictions, and finally   
   feeding the reward value back for training - the neural net in this   
   situation has only one output. Has this been done before? If so, to   
   what did it lead? And if not, is there any reason why I shouldn't try   
   it?   
      
   Thanks,   
   J. Trask   
      
   [ comp.ai is moderated.  To submit, just post and be patient, or if ]   
   [ that fails mail your article to , and ]   
   [ ask your news administrator to fix the problems with your system. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)