Heterogeneous Evolutionary Swarms with Partial Redundancy Solving Multi-objective Tasks

Consider a self-organized system of heterogeneous reconfigurable agents solving a multi-objective task. In this paper we analyze an evolutionary approach to make such a system adaptable. In principle, this system is comparable to a multi-objective genetic algorithm, however, requires asynchronous generations and decentralized evaluation- and selection processes. The primary objective of this paper is to introduce the proposed system, to provide several interesting theoretic properties and a primary experimental analysis. The heritable material (genes) compromises a parameter set that encodes an agents configuration and can be communicated between agents. We introduce partial redundancy into the system by supplying a certain number of agents with two parameter sets instead of one. These agents are denoted as redundant and are free to chose which of their two parameter sets is applied. A special focus lies on two strategies for the agents to derive a fitness value based on their property set(s) and the respective objective functions of the multi-objective task suitable for decentralized systems. A slightly more sophisticated approach with weights for each of the objectives performs just as good as a simple method where agents pick the best or respectively worst objective value. The results show that systems with low redundancy tend to lose a lot of diversity, however, redundant systems are slower in their adaptive process.