A three-neuron controller (TNC) III

A three-neuron controller is a two-layered artificial neural network (ANN) used in control problems. The activation of each neuron is above, at, or below some average or resting value. The activation of the two input neurons (lower level) is on [-1,1] and represents a normalization of the input about the resting point. The activation of the one output neuron (upper level) is on (0,1). Output values are scaled from (0,1) to their operational values by subtraction of the neuron's average value and multiplication by the appropriate ranges. The weights connecting the neurons may be positive or negative. Input activation below the average rate is negative with the magnitude representing the normalized distance below average. Hence, when its weight is multiplied by a neuron's activation, a sign change may occur and the effect of that input may appear as an excitatory or inhibitory input to the upper level node. Herein we discuss the biological plausibility of such neurons and demonstrate "stacked" TNCs (two lower level TNCs feeding their outputs as inputs to a third upper level TNC) in solving the inverted pendulum (or broom balancing) problem.