Rock-Paper-Scissors WiSARD

Abstract

This paper presents some strategies used for creating intelligent players of rock-paper-scissors using WiSARD weightless neural networks and results obtained therewith. These strategies included: (i) a new approach for encoding of the input data, (ii) three new training algorithms that allow the reclassification of the input patterns over time, (iii) a method for dealing with incomplete information in the input array, and (iv) a bluffing strategy. Experiments show that, in a tournament of intelligent agents, WiSARD-based agents were ranked among the 200 best players, one of them achieving 9th place for about three weeks.