A Parallel Network Play Backgammon that Learns to
G. Tesauro · 1989
A class of conncctionisf networks is described that has learned to play backgammon af an i~rtcr~~~rrlintr-t~~-a~l~~~t~tcc(I level. The networks were trained by back-propagation learning on a large sct of sonrpl(* pmiiions et~nlrtatcd by a human expert. In actual match play against humans and cortt~crrrio~ml cornptrtrr programs. the networks have demonstrated substantial ability to generalize on thc hmis of ierpcrt krron4rdgc ofthe game. This is possibly the most complex domain yet studied with conn~~ctiorris~ Icrrr~tirrg. Ncw trclrniqrres were needed to overcome problems due to the scale and cornplc.riy of flre tcrsk. Tlre.re hrclude techniques for intelligent design of training set examples and eficicnf coding .srhrrnrs. and procedures for escaping from local minima. We suggest how these ~cclrniqtres nrigltt bc used in applications of network learning to general large-scale, diflcrilt real-worhi prob/cm domains.