Abstract: | Probabilistic Incremental Program Evolution (PIPE) iteratively generates populations of programs from a probability distribution overall possible programs, given a function and terminal set, evaluates the population and uses the best program to update the probability distribution. In that way the search through the space of possible solutions is guided towards the optimum. Because fewer operations are needed, PIPE is computationally cheaper than comparable techniques like Genetic Programming. Experiments are reported that show the ability of PIPE to evolve and co-evolve controllers for simulated Khepera robots in obstacle avoidance and pursuit/evasion scenarios.
|