Intelligent Mobile Robot Group
National Taiwan University
June 1, 1997
Submitted to HICSS-31
Intelligent agents who participate in free-forming collaborations choose strategies to maximize their effectiveness in achieving goals. In a heterogeneous agent society where each has no knowledge of the logic behind other's motives or real intentions, it is difficult to conceive how an algorithm would provide an effective collaboration strategy. We suggest a set of general, quantitative criteria for detecting collaboration situations leading to the duoagent collaboration dilemma (DCD). We show that DCD is a widely applicable collaboration scenario and it can take advantage of a class of problems social scientists have studied extensively. Lastly, we provide steps to developing practical computable strategies for agents to avoid and resolve problems in a subclass of DCD.
We have identified the requirement for multiagent collaboration scenarios where strategies for the prisoner's dilemma game (PDG) can apply. For this domain, called the discernable duoagent collaboration dilemma (DDCD), we have developed mechanisms to program intelligent agents to implement any computable game theoretic strategy found for the PDG, all without violating the assumption of computational imprecision. The agents base their decisions purely on profits or deficits for other agents, which are recorded in binary history by some neutral referral agency. This is in contrast against game theoretic results which depend on the decisions of the agents that are difficult to recognize by other agents in practical terms.
We have also, on a general basis, introduced the formal notion of collaboration efforts, effectiveness, profits, and profit share. Based on these structures we can define payoffs that are linear in nature, which are simple to evaluate arithmetically and have error ranges bound within a multiple of the sampling precision. This may serve as a promising framework for future investigations.
Individual uncompressed files:
Created: June 8, 1997