Methods for evolving robust distributed robot control software: coevolutionary and single population techniques

Previous work on evolving distributed control software for modular robots has resulted in solutions that do not generalize well to unseen test cases. In this work, we seek general solutions to an entire space of test cases. Each test case is a specific world configuration with a passage through which the modular robot must move. The space of test cases is extremely large, so a given training set can only be a sparse sample of this space. We look at several approaches for dealing with the problem of determining an effective training set: using a fixed set throughout a run, sampling randomly at each generation, and using coevolutionary approaches to evolve a population of test worlds. For this problem, random sampling outperformed the fixed sampling technique and did at least as well as the coevolutionary techniques we considered