Application of the BOXES methodology to continuous systems

This paper details the application of an AI technique known as the BOXES method to continuous, real-time, real-world processes. The proposed controller requires no mathematical pre-definition of the process under control, and operates in a truly black-box fashion. The BOXES algorithm has two distinct operational modes viz: react and reflect. In the react mode, binary control decisions are obtained by reading the elements of a control matrix indexed by a unique numerical identifier computed from the current system variable values. In the reflect mode, data values from the run just ended are aggregated with all runs to date and the control matrix updated using statistical inferencing. In this application, the BOXES algorithm is made to augment the integral and derivative terms of a PID schema. The automaton is set the task of learning to minimize the divergence of the output variable from an applied demand stimulus. The paper describes the BOXES philosophy and an adaptation of the method to continuous systems such as are found in process-control situations. The method hinges on the expression of performance merit using time-variant data and the attribution of penalty and reward to only those states contributing to the merit in any given run.