No abstract

This paper for the special session on Adaptive Critic Design Methods at the SMC '97 Conference describes a modification to the (to date) usual procedures reported for training the Critic and Action neural networks in the Dual Heuristic Programming (DHP) method [7]-[12]. This modification entails updating both the Critic and the Action networks each computational cycle, rather than only one at a time. The distinction lies in the introduction of a (real) second copy of the Critic network whose weights are adjusted less often (once per "epoch", where the epoch is defined to comprise some number N>1 computational cycles), and the "desired value" for training the other Critic is obtained from this Critic-Copy.In a previous publication [4], the proposed modified training strategy was demonstrated on the well-known pole-cart controller problem. In that paper, the full 6 dimensional state vector was input to the Critic and Action NNs, however, the utility function only involved pole angle, not distance along the track (x). For the first set of results presented here, the 3 states associated with the x variable were eliminated from the inputs to the NNs, keeping the same utility function previously defined. This resulted in improved learning and controller performance. From this point, the method is applied to two additional problems, each of increasing complexity: for the first, an x-related term is added to the utility function for the polecart problem, and simultaneously, the x-related states were added back in to the NNs (i.e., increase number of state variables used from 3 to 6); the second relates to steering a vehicle with independent drive motors on each wheel. The problem contexts and experimental results are provided. BACKGROUNDDual Heuristic Programming (DHP) is a neural network approach to solving the Bellman equation [12]. The idea is to maximize a specified (secondary) utility function:The term is a discount factor ( ) and is the primary utility function, defined by the user for the specific application context. A useful identity:(2) In this paper, is assumed to be 1, and the usual meth-ods of discretizing continuous models of plants is used. For DHP, at least two neural nets are needed, one for the actionNN functioning as the controller, and one for the criticNN used to train the actionNN. A third NN could be trained to copy the plant if an analytical description (model) of the plant is not available. R(t) [dimension n] is the state of the plant at time t. The control signal u(t) [dimension a] is generated in response to the input R(t) by the actionNN. The signal u(t) is then asserted to the plant. As a result of this, the plant changes its state to R(t+1). The criticNN's role is to assist in designing a controller (actionNN) that is "good" relative to minimizing the specified cost function U(R(t),u(t)), which is designed to expresses the objective of the control application. In the DHP method, the criticNN estimates the gradient of J(t) with respect to R(t); the letter λ is used as a short-hand no...