PERCEPTIVE PARTICLE SWARM OPTIMISATION: AN INVESTIGATION Boonserm Kaewkamnerdpong and Peter J. Bentley Department of Computer Science, University College London, UK {b.kaewkamnerdpong, p.bentley}@cs.ucl.ac.uk ABSTRACT Conventional particle swarm optimisation relies on exchanging information through social interaction among individuals. However for real-world problems involving control of physical agents (i.e., robot control), such detailed social interaction is not always possible. Recently, the Perceptive Particle Swarm Optimisation (PPSO) algorithm was proposed to mimic behaviours of social animals more closely through both social interaction and environmental interaction for applications such as robot control. In this study, we investigate the PPSO algorithm on complex function optimisation problems and its ability to cope with noisy environments. 1. INTRODUCTION In particle swarm optimisation, all individuals in the swarm have the same behaviours and characteristics. It is assumed that information on the position and performance of particles can be exchanged during social interaction among particles in the neighbourhood. Importantly, the conventional particle swarm optimisation relies on social interaction among particles through exchanging detailed information on position and performance. However, in the physical world, this type of complex communication is not always possible. Global communication may be impossible amongst swarm of robots. Indeed, it is common for robots to have no idea of their own performance at a given location and thus there may be little direct information that one individual can pass on to its companions. Insects must cope with similar problems. Termites do not build their mounds by talking to each other and telling each other where to deposit material. Instead, they perceive each other, and they perceive their environment, and their complex behaviour emerges as a result of those perceptions. There is no concept of communication, only interaction. Social interaction and environmental interaction (stigmergy) enables termites to build highly complex structures without direct communication [1, 2]. This work focuses on the use of swarm intelligence for physical applications, where these kinds of severe communication restrictions are common. In order to imitate the physical collective intelligence in social insects, we previously proposed the Perceptive Particle Swarm Optimisation (PPSO) algorithm, which adds an extra dimension to the search space and enables both social interaction and environmental interaction by allowing a finite perception range for each individual [3]. In this study, we investigate the performance of the PPSO algorithm on complex function optimization problems and its ability to cope with noisy environment. The conventional particle swarm optimisation and its modifications including the PPSO algorithm are described in section 2. The PPSO algorithm is discussed in comparison to conventional particle swarm optimisation. In section 3, the aim of the investigation and the methodology are discussed. Section 4 describes experiments to investigate the performance of PPSO and conventional particle swarm optimisation according to the methodology. A discussion of the experimental results is provided in section 5. 2. BACKGROUND Conventional Particle Swarm Optimisation The particle swarm optimisation algorithm was introduced by Kennedy and Eberhart in 1995 [4]. The algorithm consists of a swarm of particles flying through the search space. Each individual i in the swarm contains parameters for position x i and velocity v i , where x i ∈ R n , v i ∈ R n while n is the dimension of the search space. The position of each particle represents a potential solution to the optimisation problem. The dynamics of the swarm are governed by a set of rules that modify the velocity of each particle according to the experience of the particle and that of its neighbours depending on the social network structure within the swarm as shown in equation 1. By adding a velocity vector to the current position, the position of each particle is modified. As the particles move around the space, different fitness values are given to the particles at different locations according to how the current positions of particles satisfy the objective. At each iteration, each particle keeps track of its personal best position, pbest. Depending on the social network structure of the swarm, the global best position, gbest, and/or the local best position, lbest, is used to influence the swarm dynamic. After a number of iterations, the particles will eventually cluster around the area where fittest solutions are. v i (t+1) = w.v i (t) + c 1 r 1 (x pbest,i - x i (t)) + c 2 r 2 (x gbest - x i (t)) (1) x i (t+1) = x i (t) + v i (t+1) The particle swarm optimisation algorithm has been successfully employed to solve a range of optimisation problems including electric power systems [5], music [6], human tremor analysis [7], image classification [8], logic