Novelty-driven Particle Swarm Optimization

Particle Swarm Optimization (PSO) is a well-known population-based optimization algorithm. Most often it is applied to optimize objective-based fitness functions that reward progress towards a desired objective or behavior. As a result, search increasingly focuses on higher-fitness areas. However, in problems with many local optima, such focus often leads to premature convergence that precludes reaching the intended objective. To remedy this problem in certain types of domains, this paper introduces Novelty-driven Particle Swarm Optimization (NdPSO), which is motivated by the novelty search algorithm in evolutionary computation. In this method particles are driven only towards instances significantly different from those found before. By ignoring the objective this way, NdPSO can circumvent the problem of deceptive local optima. Because novelty search has previously shown potential for solving tasks in genetic programming, this paper implements NdPSO as an extension of the grammatical swarm method, which combines PSO with genetic programming. The resulting NdPSO implementation is tested in three different domains representative of those in which it might provide advantage over objective-driven PSO. That is, deceptive domains in which it is easy to derive a meaningful high-level description of novel behavior. In each of the tested domains NdPSO outperforms both objective-based PSO and random-search, demonstrating its promise as a tool for solving deceptive problems.