Predatory sequence learning for synthetic characters

Abstract

The process of mammalian predatory sequence development offers a number of insights relevant to the goal of designing synthetic characters that can quickly and easily learn complicated and interesting behavior. We propose a number of principles for designing such learning systems, inspired by a targeted review of animal developmental phenomena, with particular emphasis on the development of predatory behavior in certain felid and canid species. We describe the implementation of a few of these principles as an extension to a popular algorithm for learning in autonomous systems called hierarchical Q-learning. In this new approach, the agent starts out with only one skill, and then new skills are added one at a time to its available repertoire as time passes. The agent is motivated to experiment thoroughly with each new skill as it is introduced. Simulation results are presented which empirically demonstrate the advantages of this new algorithm for the speed and effectiveness of the learning process.