Sensor-driven associative random optimization (SARO) for control

A novel mechanism for adaptation of synaptic weights of a neural control net, using a separate `training net´ with sensor-driven associative random optimization (SARO), is proposed. SAROnet adjusts all weights of the control net in parallel, as defined by a scalar `Critic´, which evaluates the performance error of the controlled system. The sensory input for SAROnet may be composed of the control net input and the output of the controlled system. The learning performance of SARO for global and stepwise error minimization was successfully tested with simulations for several n-bit parity tasks and the 8-3-8 encoding problem. Speed of convergence for n-bit parity tasks was found to be considerably higher as compared to error-backpropagation or a more recent random optimization method. Based on these encouraging benchmark tests, SAROnet will be applied to control tasks