ON THE OPTIMAL-CONTROL OF BEHAVIOR - A STOCHASTIC PERSPECTIVE

Evolution is a closed stochastic optimisation process driven by the in teraction between behaviour and environment towards local maxima in fi tness. It is inferred that nervous systems are selected to provide opt imal control of behaviour (the 'assumption of optimality'), such that for some behaviours, the expectation of future hazards to survival are minimised. This is illustrated by goal-directed saccades in which min imising total flight-time of primary and secondary movements provides a better fit to observations than simply minimising the error of the p rimary movement. This optimisation is extended to intra-movement traje ctories, where low-bandwidth (smooth) velocity profiles provide a more satisfactory description of observations than simple bang-bang contro l. Since minimum-time behaviours cannot be controlled by error feedbac k, it is concluded that the cerebellum must be executing a real-time u nreferenced optimisation process. This requires explorative as well as exploitative behaviour. Stochastic gradient descent is discussed as a possible means by which the cerebellum may optimise behaviour. (C) 19 98 Elsevier Science B.V. All rights reserved.