A distributed learning algorithm for particle systems

Particle systems, a type of swarm intelligence system, have repeatedly been shown to be capable of more general problem solving than just collective movements. The emerging, self-organizing behavior of such systems leads to collective behavior that tends to be far more complex than that of their parts, yet at the same time their self-organizing nature typically makes their behavior difficult to predict and control. In previous work, we introduced a set of mechanisms to guide the self-organizing process, allowing the system designer to exert a form of high-level control over a self-organizing system. Here we extend these past results by incorporating a "pollen-based" distributed learning algorithm that increases the capabilities of a team of cooperating agents in pursuit of a global goal, while still retaining most of the simplicity of particle systems. To demonstrate these ideas, we use a dynamic logistics problem that combines the coordination and cooperation issues of collective transport with the global optimization difficulties of routing and shop-floor scheduling problems. The results show that this form of dynamic distributed learning enables a particle system to function effectively and to adapt to changing conditions relatively quickly. Further, the results suggest that the combination of distributed learning with collective movements provides an additional advantage that significantly affects system-wide behavior.