Learning by On-Line State Space Refinement

An approach for automatic control using on-line learning is presented. This approach uses a set of adaptive elements, each of which corresponds to a (hyperspherical) region of state space. The state space partition is initially coarse. An adaptive element is activated when the state enters its region, allowing the element to attempt to optimize the response within this region. The region size is periodically reduced to permit the element to obtain a more accurate suboptimal response. New adaptive elements are generated to maintain coverage of state space, with characteristics drawn from neighboring regions. During on-line operation the resulting state space partition undergoes successive refinement, with learning (and memory requirements) concentrated in the most frequently-arising areas, and with "generalization" occurring over successively smaller areas.