We view communication as action aimed at increasing the efficiency of interaction among agents. Thus, we postulate that a speaker design a speech act so as to maximally increase the benefit it obtains as the result of the interaction. Our framework, based on Recursive Modeling Method (RMM), consists of a representation of an epistemic state of an agent engaged in an interaction, and includes the agent's preferences, abilities and beliefs about the world, as well as the beliefs the agent has about the other agents, the beliefs the other agents have, and so on.
Given the RMM representation, a pragmatic meaning of a speech act can be defined as a transformation it induces on the epistemic state of an agent. This transformation leads to a change in the quality of the interaction, expressed in terms of the benefit to the agent. We propose that the rational communicative behavior results from a speaker choosing to perform the speech act that maximizes the expected increase in the quality of the interaction.
Each of the two defense units can launch three interceptors and are faced with an attack by seven missiles named as A, B, C, D, E, F, and G respectively. In these example runs, missile warhead sizes for A..F are 470, 410, 350, 370, 420, 450, 430. Each case below is executed over 100 times, and the quality of the coordination acheived is measured as the combined expected tonnage of the missiles that penetrated the defense and damaged the protected territory (since each interceptor has an interception probability of less than one, the residual probability that the missile has not been destroyed is multiplied by its size and included in this measure).
We allowed for one-way communication at a time between defense units. And then, each battery was assumed to have a choice of the following communicative behaviors: "No communication," "I'll intercept missile A." through "I'll intercept missile G," "I have both long- and short-range interceptors," "I have only long-range interceptors," or "I'm incapacitated."
Among the various episodes we ran we will consider two illustrative examples to examine the coordination achieved by RMM and human team in more detail. In these examples each defense unit was fully functional and has both long- and short-range interceptors. The two defense agents by RMM or human team, labeled "1" and "2", attempt to shoot down the seven incoming missiles when communication was, and was not, available.
Please allow about 2 min. before the initial screen after loading demo.
Scenario 1 The
targets are scattered over the defense agents 1 and 2.
Scenario 2 The
targets are clustered in front of the defense agent 1. In this case,
communication is more critical because it is more likely that targets
could be intercepted redundantly without communication.
The experimental results present the average total expected damage by
RMM agents after 100 trials and by human agents after 20 trials. The
team of two RMM agents, when faced with a random arrangement of the
attacking missles, allowed, on the average, the combined tonnage of
717.01 (s=110.71) to penetrate the defense, while the human team
scored 800.45 (s=147.69). The total expected damage of two RMM agents
with communication was, on the average, 652.33 (s=58.97), while the
damage suffered by the human team was 710.20 (s=100.92)
The results are intuitive: as expected, communication improves the coordinated performance achieved by the teams. It is interesting to see the differences between the communicative behaviors exhibited by RMM and human agents. While human communicative behaviors were often similar to those selected by the RMM agents, there are telling differences that, in our experimental runs, allowed the RMM team to achieve a slightly better performance.
This research has been sponsored by the Office of Naval Research Artificial Intelligence Program under contract N00014-95-1-0775. Here are the papers describing our approach in more detail:
For more information, please email us.
