Co-adaptive learning classifier systems based on coevolution within dyna architecture
Chungynan Huang, Chuen–Tsai Sun · 2004
Learning classifier systems are a model for problem-independent and adaptive machine learning. As with evolutionary computations, parameter settings determine whether learning classifier systems can generate optimal solutions and whether it can do so efficiently. The authors propose a co-adaptive approach to controlling parameters for coevolution-based learning classifier systems. By taking advantage of the on-line incremental learning capability of such systems, solutions can be produced that completely cover a target problem. The system combines the advantages of both adaptive and self-adaptive parameter-control approaches. Using a coevolution model means that two learning classifier systems can operate in parallel, to simultaneously solve target and parameter-setting problems. Furthermore, the approach needs very little time to become efficient in terms of latent learning, since it only requires small amounts of information on performance metrics during early run-time stages. Our experimental results show that the proposed system outperforms comparable models regardless of a problem's stationary/non-stationary status.