FEP learning algorithm: application to direct self-learning control

One significant problem with neural control is that of credit assignment, that is, how should errors in the plant output be used to modify the controller, since the plant is interposed between the controller output and the “scored” output. While the controller takes as inputs, the plant outputs, then we can propagate forward the error through the controller network, and then we update its weights. This is the basic principle of the feedforward error propagation (FEP) learning algorithm developed in the paper. This new algorithm does not need a second network to train the controller. This avoids the extra network uncertainty and greatly simplifies the computation complexity and thus, makes it suitable for online learning. The FEP algorithm is used to design a direct self-learning control system for the inverted pendulum and its performance is compared with that of backpropagation based self-learning control