Generality Vs. speed of convergence in the cart-pole balancer

This paper compares the speed of convergence to an optimal solution of four controllers for the problem of balancing a pole on a cart. We demonstrate that controllers whose design is tailored specifically to the cart-pole problem (i.e. less general) converge more rapidly to an optimal solution. However, the architectures and learning algorithms for those networks may not perform well for more general tasks. The four controllers, ordered from the least general to the most general, are the Perceptron, the Associative Search Element [1], Jordan's approach [3], and Prejudicial Search applied to the ASE architecture. Two of the above neural networks, the Perceptron and Prejudicial Search [2] are new methods for solving this problem. The perceptron is a simple two input neuron (summing unit) with 2 weights and a step function output. The Prejudicial Search is a method for biasing the search of possible solutions. It guarantees convergence, but allows the search to be biased by heuristics or information about the problem. In this paper, it is combined with the ASE architecture. However, the Prejudicial Search technique can be combined with any architecture and learning algorithm, extending their ability to handle a more general class of problems.