A new basis for state-space learning systems and a successful implementation

摘要

A new basis for state-space learning systems is described which centers on a performance measure localized in feature space. Although a parameterized linear evaluation function is its end product the iterative implementation examined has elements both of adaptive control (parameter optimization) and of pattern recognition (pattern formation and feature selection). The system has repeatedly generated an evaluation function which solves all of a random set of fifteen puzzle instances. Despite the absence of any objective function the parameter vector is locally optimal, a result not previously attained in any way.