| Popis: |
The use of a global fitness function shows the possibility of training multirobot controllers with limited supervision to perform self-organized task decomposition. A global fitness function encourages solutions that improve system performance without explicitly biasing for a particular task decomposition strategy. This is advantageous for use with multirobot controllers, where it is often easier to define the global goals over required local coordination behaviours that require task specific information. In addition, such an approach facilitates discovery of novel behaviours that otherwise may be overlooked by a human designer. Such novel behaviours include use of `bucket brigade' behaviours for the multirobot resource gathering task, `error correction' for the tiling formation task and incremental merging of the object piles for the heap formation task. For the walking gait task, use of a global fitness function facilitated discovery of the local leg behaviours and resultant global gait coordination without the need for multistaged evolution. While some of the behaviours from evolved controllers may display novel attributes, quantitatively interpreting such solutions remains difficult. This issue is of added importance because we are interested in how a set of local behaviours give rise to emergent global behaviours for multirobot control. As Dawkins (1986) points out, successful evolutionary solutions are not evolved per se to ease our burden of understanding such solutions. Biologically plausible techniques such as the ability to track energy consumption of the individual robots and bias for minimization of energy consumption within the fitness function also does not guarantee more decipherable solutions nor improved evolutionary training performance than without selecting for energy efficient controllers. However, such implicit techniques may well reduce any redundant behaviours that may be energetically wasteful. Alternatively, desired attributes could be explicitly encouraged through shaping, helping to simplify deciphering such solutions but this implies dealing with the problem of greater supervision for multirobot control and how best to perform decomposition of the task a priori. Comparison of the different evolutionary techniques discussed in this chapter shows that by exploiting hierarchical modularity and regulatory functionality, controllers can overcome tractability concerns. Simple CA lookup table controllers are shown to be suitable for |
