| Popis: |
This dissertation focuses on the evolution of Continuous Time Recurrent Neural Networks (CTRNNs) as controllers for control systems. Existing research suggests that the process of neutral drift can greatly benefit evolution for problems whose fitness landscapes contain large-scale neutral networks. CTRNNs are known to be highly degenerate, providing a possible source of large-scale landscape neutrality, and existing research suggests that neutral drift benefits the evolution of simple CTRNNs. However, there has been no in-depth examination of the effects of neutral drift on complex CTRNN controllers, especially in the presence of noisy fitness evaluation. To address this problem, this dissertation presents an analysis of the effect of neutral drift on the evolution of a complex CTRNN locomotion controller for a simulated hexapod robot in the presence of noisy fitness evaluations. In particular, two stochastic hill-climber-based EAs are examined and compared, one that does not engage in neutral drift, and one that does. The experimental results show that while neutral drift provides a significant advantage early in the evolutionary process, the later effects of noisy fitness evaluations seriously degrades the utility of neutral drift, and overall, there is no significant difference between the non-drifting and drifting EAs. These results provide evidence that large-scale neutral networks do exist in complex CTRNN fitness landscapes and highlight the important role that noisy fitness evaluations play in influencing evolutionary performance. |
