Cellular Automata-Based Scheduling: A New Approach to Improve Generalization Ability of Evolved Rules

This paper presents a cellular automata-based algorithm designed to schedule tasks for parallel processors. In the learning phase, a genetic algorithm is used to discover cellular automata (CA) rules able to solve an instance of a multiprocessor scheduling problem. In the normal operating phase, the discovered rules are applied to find optimal or suboptimal solutions to other scheduling problem instances. A new approach to the learning phase is presented here, called Joint Evolution. The results obtained have shown evolved rules with a better generalization ability when they are applied to small variations of the problem used as base for the evolution.