sammut@qut.edu.au (David Sammut) (04/19/91)
I am currently working at implementing the broom balancing problem using an adaptive heuristic critic much like in the paper by Barto, Sutton and Anderson called "Neuronlike Adaptive Elements That Can Solve Difficult Learning Control Problems". However, I am not implementing the decoder as shown in the paper (that decodes the input vector into 162 binary values of which only one has been set) but attempting to use this input vector as straight input into the network. My reason for atttempting this is because the book in which I got my original AHC algorithm from said that it was capable of accepting analog inputs (e.g. the component values of the vectors should do). Is it possible to solve the problem in this way? It seems that it begins to learn (the time before failure improves on the first 20 or so training runs) but then declines slowly(forgets?) so that the pole is eventually failing in near minimum time. Any thoughts anyone ? I'd be extremely grateful for any helpful suggestions. (The book that I got the algorithm from was called Artificial Neural Systems - Foundations, Paradigms, Applications, and Implementations (Patrick K Simpson). It is quite a good book for an overview of the various paradigms that have been developed.) David Sammut +-----------------------------------------------------------------------+ | David Sammut | | | School of Computing Science | | | Faculty of Information Technology | email: sammut@qut.edu.au | | Queensland University of Technology | | | Brisbane, Queensland, Australia | | +-----------------------------------------------------------------------+