Towards agent Swarm Optimization
Dario Schor, Witold Kinsner · 2011
This paper examines particles in Particle Swarm Optimization (PSO) in terms of situated agents to improve control of the long term behaviour of particles as required for cognitive machines. In PSO, the particles lack goals and temporal knowledge for personal/global best solutions, thus many of the decisions for advancing the position of the particle diverge away from the solution causing the algorithm to take longer as the particles return to their intended path. This paper proposes novel modifications to the standard PSO algorithm to incorporate self-adjusting temporal knowledge to help guide the particles towards the optimal solution faster. The temporal knowledge is incorporated as a weighted sum of L previous steps taken by the particle, where L is automatically adjusted to maintain a certain multiscale measure that satisfies a balance between exploration and seeking the goal. Additional improvements based on the dynamics of the particle's behaviour are described that would allow for real-time predicting of parameters.