Negotiated decommitment in a collaborative agent environment
Michele Van Dyne, Costas Tsatsoulis · 2003
Decommitment of agents has been studied in the context of self-interested agent societies, primarily using a leveled commitment protocol in which agents committing a breach of contract are penalized in various ways. Limited research has investigated decommitment in collaborative agent populations. This research investigates decommitment in a collaborative and benevolent agent society, in which agents make decisions and act in a manner which maximizes the global benefit. In this environment, decommitment takes on a different nuance than much of what has been presented in the literature, in that collaborative agents can be expected to behave so as not to cause harm to the agent society as a whole, even in decisions not to honor commitments. When an agent perceives that decommitting to a previous commitment has the potential for increasing its local utility, the agent first attempts to negotiate the terms of the decommitment with those affected. In the context of collaborative agents, it was hypothesized that allowing agents to decommit would have an overall beneficial effect on attaining system goals, however, decommitment would require additional overhead as far as information required and communication resources. Thus, there would be a trade-off between the resource constraints placed on the system versus the beneficial effect of decommitment. Negotiated decommitment was expected to provide more of a beneficial effect than unilateral decommitment. Testing of the theory and experimentation was performed in the context of the Autonomous Negotiating Teams (ANTS) system. Agents are based on the BDI agent architecture model and have the characteristics that they are rational, collaborative, benevolent, homogeneous and multitasking, and they are capable of negotiation and planning time-dependent activities. The operating environment is both time-bounded and resource constrained. Results indicated that negotiated decommitment performed as hypothesized, that is, overall performance improved. Graceful degradation of performance was demonstrated in all conditions as the constraints on the system increased. Under all conditions, agents demonstrated rationality in decision making and both negotiated decommitment and unilateral decommitment provided a higher locally assessed utility than did the baseline condition.