Abstract: | Credit assignment is a major problem in making tractable the learning process in all but the simplest of environments. Even simple arcade-style games have proven to be problematic. This project uses a GP-based incremental evolution in an attempt to evolve a competent game agent for an "Asteroids"-style universe. Problems encountered during the course of the project highlight the difficulties of such a venture even with an incremental approach.
Key features of this project include:
- The use of incremental evolution to tackle the evolution of the genome in smaller, more tractable chunks, hopefully reducing the credit assignment problem.
- The development of an unorthodox, highly structured multi-tree genome architecture for the GP, intended to facilitate better context-preserving crossover.
- The development of a domain-specific control architecture, featuring a heterarchical votes and vetoes decision system.
- The employment of competitive rather than objective fitness functions to co-evolve behaviours at certain stages of the incremental competence acquisition process.
The project had areas of success, but the overall quality of results suggests that a GP approach is only likely to work well in this area with certain suggested improvements.
|