Nesting, Safety, Layering, and Autonomy: A Reorganizable Multiagent Cerebellar Architecture for Intelligent Control with Application in Legged Robot Locomotion and Gymnastics

摘要

The learning and control space of real-world au- tonomous agents are often many-dimensional, growing, and un- bounded in nature. Such agents exhibit adaptive, incremental, exploratory, and sometimes explosive learning behaviors. Learn- ing in adaptive neurofuzzy control, however, is often referred to as global training with a large set of random examples and a very low learning rate. This type of controller is not reorganizable; it cannot explain exploratory learning behaviors as exhibited by human and animal species. A theory of coordinated computational intelligence (CCI) is proposed in this paper which leads to a reorganizable multiagent cerebellar architecture for intelligent control. The architecture is based on the hypotheses that 1) a cerebellar system consists of a school of relatively simple and cognitively identifiable semiautonomous neurofuzzy agents; 2) autonomous control is the result of cerebellar agent fine-tuning and coordination rather than complicate computation; and 3) learning is accomplished via individual cerebellar agent learning and coordinated discovery in a learning-tuning-brainstorming process. Agent-oriented decomposition and coordination algo- rithms are introduced; necessary and sufficient conditions are established for cerebellar agent discovery and common sense cerebellar motion law discovery. Nesting, safety, layering, and autonomy—four principles are analytically formulated for the reorganization of neurofuzzy agents. The four principles ex- tend Saridis' principle of increasing intelligence with decreasing precision for hierarchical control to multiagent neurofuzzy con- trol, and bridge a gap between neurofuzzy control and PID control, adaptive learning and exploratory learning, numerical learning and symbolic learning. Basic ideas are illustrated in legged locomotion and gymnastics. It is shown that, with agent- oriented decomposition, a single near-miss example can enable the cerebellar agents of a 3-link or 4-link uniped simulation to learn gymnastic jumps; a small number of fine-tuned agents can form a kernel community which can discover common sense cerebellar motion laws; and a kernel governed by the laws can grow with a geometrical learning rate. Implications of this work to human and animal locomotion control are discussed. Potential applications and extensions of the findings are outlined.