Synergy in cooperating agents: designing manipulators from task specifications

Abstract

This thesis considers two design problems. The first, called the modular manipulator design problem, is to design a new manipulator from a bin of parts, each time a new task is to be performed. Tasks are specified by sets of kinematic and dynamic constraints. The design of the manipulator must meet these constraints while providing a good trade-off among the conflicting criteria of minimum weight, deflection and maximum dexterity. The modular manipulator design problem is exceedingly difficult. Available algorithms can only tackle parts of it, typically, focusing on one or two of the five criteria while neglecting the rest. This situation leads to the second problem; organizing algorithms into a team so that they can cooperate and thereby solve problems well beyond their individual capabilities. We use an existing organizational structure, an A-team, in which autonomous agents cooperate by working iteratively, asynchronously and in parallel to improve a population of designs in a common memory. We primarily use two kinds of agents; modification operators that improve designs by using simple qualitative knowledge and destroyers that delete designs to keep the population in check. To ensure quick continuous improvement of the population, we use two strategies: (1) Modification operators cooperate using an extension of means-ends analysis; they work mainly on designs that they can improve. (2) Destroyers selectively delete designs that are far away from the Pareto set. Together these strategies herd designs towards the Pareto set. This straightforward organization of independent agents, cooperating using these simple strategies, provides quick comprehensive solutions to the problem of designing modular robots from task specifications and makes an RMMS realizable. Since all agents in an A-team are independent, and cooperate only by sharing results, addition and deletion of agents is trivial and the organization is truly modular. An A-team thus robustly combines diverse heuristics in a modular fashion to solve complex problems, even though these heuristics (a) each address only a subset of the criteria and (b) may be in conflict. It achieves results better than any single agent can.