Demetri Psaltis and Christof Koch
Computation and Neural Systems Program
California Institute of Technology
Designing intelligent systems that can monitor and interact with complex, variable, and poorly modeled environments remains a challenge. This is particularly true for systems that need to be controlled in real-time, such as autonomous robots, automated buildings, and traffic control in metropolitan areas. We describe an approach that sacrifices the time-consuming (and, for many physical systems, ill-defined) goal of searching for the global optimum in favor of a locally optimal solution in a small, restricted subset of the system. This ``region of interest'' is determined in real-time as the best representation of the system status given limited computational resources, and changes as the system and the environment evolve. The organization of our model is reminiscent of the cognitive architecture of the primate brain and, in particular, to the function of consciousness/awareness as proposed by Crick and Koch.
We exemplify our strategy with two examples. (i) The implementation of a two-player competitive video game of "Desert Survival" and (ii) in the context of a well-known problem in computational complexity. This involves the performance of "Match Fit", a novel, on-line bin packing algorithm, which can interpolate smoothly from the "Next Fit" to "Best Fit" algorithms. It is based on a heuristic which packs multiple blocks at once. The performance of this O(n) on-line algorithm can be better than that of the Best Fit algorithm. On large sample problems, the new algorithm runs about an order of magnitude slower than Next Fit, and about two orders of magnitude faster than Best Fit. It can be tuned for optimality in performance by adjusting parameters which set its "working memory" usage, and exhibits a sharp threshold in this optimal parameter space as time constraint is varied. These optimality concerns provide a testbed for applying certain key features of the cognitive architecture of the primate forebrain, working memory and an attentional selection process, to algorithms.