Emergence of Collective Intelligence in Stochastic Local Search-and- Optimization Systems

Two new stochastic search methods are introduced as prototypic examples showing how collective intelligence may emerge in a system of locally interacting units. They share the property of being theoretically understandable and computationally tractable, a quite “rare” feature. The first search method, based on multiresolution search algorithms, can be typically implemented under the form of search agents. The method is appropriate if the target element(s) is located in high-average fitness regions of the search space. The search may be improved by introducing some interaction between the search agents. As the agents search preferentially in high-average fitness regions, there is a correlation between the number of agents in a region of the search space and the local average fitness in that region. It is therefore natural to introduce some extra sampling when several agents are in the same neighborhood. The theoretical framework of multiresolution analysis and wavelet theory permits to give a precise description of the above strategy and to define simple conditions guarantying that the search with interacting agents is better than a search with a single agent. The second example shows how a satisfiability problem (3-SAT) can be solved by an ensemble of small computing units working in parallel. The search uses a number of noisy integrate-and-fire neurons as local optimizer. The satisfiability problem is coded so that if a solution does exist then the integrate-and-fire system is in a ground state. The resulting algorithm is new, easily scalable and is better than existing stochastic algorithms for random nonstationary 3-SAT problems. Under some particular conditions, the algorithm reduces to RwalkSAT, a local stochastic search algorithm whose properties in conjunction to the random 3-SAT have been recently explained theoretically. The emergence of intelligence in the noisy integrate-and-fire neurons can therefore be at least qualitatively understood on the basis of these results.