 |
Josh Bongard |
|
Associate Professor |
|
329 Votey Hall, Burlington, VT 05405 |
|
Tel: (802) 656-4665; Fax: (802) 656-0696 |
|
Curriculum Vitae: PDF |
|
| New | Research | Teaching | Media | Publications | The Zoo | PhD Thesis | MSc Thesis | Misc |
| MSc Thesis, University of Sussex |
| [PS] |
Evolving Heterogeneity: Implications for Agent-Based Systems and Collective Problem Solving This report investigates the dynamics of agent populations evolved to perform some collective task, and in which the heterogeneity of the group behaviours is under evolutionary control. Two task domains are studied: for the first task, both an optimal homogeneous and an optimal heterogeneous solution exist; the optimal behaviour for the second task is unknown. Both homogeneous and heterogeneous solutions are evolved for the two tasks using a population-level evoltuionary algorithm called the Legion system. A new metric of heterogeneity is introduced, which measures the heterogeneity of any evolved group behaviour. It was found that among evolved populations with similar fitness values, behaviours with a high heterogeneity measure had simpler control architectures than behaviours with a low heterogeneity measure. Similarly, by evolving collective behaviours which are generated by neural networks, it is shown that smaller networks produce more heterogeneous populations than relatively larger networks. It is hypothesized that the smaller size of heterogeneous control architectures is a result of specialization. This hypothesis is supported by results from the food foraging problem in simulated ant colonies: heterogeneous solutions tend to have less unused control structure than similarly fit homogeneous solutions. Moreover, increased specialization is found in populations controller by smaller networks. Specialization is analyzed in detail among a set of four evolved group behaviours. Finally, the implications of heterogeneity for evolutionary robotics and other evolutionary computation domains are discussed. |