The purpose of this presentation is to review and clarify some of the foundational concepts and mechanisms that underlie parallel problem solving in nature. A problem can be conceived as a tension between the present, "unfit" state and some fit state in which the tension would be relaxed [2]. Formulated in terms of dynamic systems, the solution is then a fitness peak, a potential valley, or most generally an attractor in the state space of the system under consideration. Solving the problem means finding a path that leads from the present state to such an attractor state. This spontaneous descent of a system into an attractor is equivalent to the self-organization of the components or agents in the system, meaning that the agents mutually adapt so as to achieve a stable interaction pattern. The interaction between agents can be conceived as a propagation of challenges: a challenge is a state of tension that incites an agent to act so as to reduce the tension. That action, however, typically creates a new challenge for one or more neighboring agents, who act in turn, thus creating yet further challenges. The different actions take place in parallel, producing a "wave" of activity that propagates across the environment. Because of the general relaxation dynamics, this activity eventually settles in an attractor. The stability of the resulting global configuration means that the different agents have now "coordinated" their actions into a synergetic pattern: a global "order" has emerged out of local interactions [1]. Such self-organization and "natural problem solving" are therefore in essence equivalent. Two mechanisms facilitate this process: (1) order from noise [4] notes that injecting random variation accelerates the exploration of the state space, and thus the discovery of deep attractors; (2) stigmergy means that agents leave traces of their action in a shared medium. These traces challenge other agents to build further on the activity. They function like a collective memory and communication medium that facilitates coordination without requiring either top-down control or direct agent-to-agent communication [3].
展开▼