Most inter-agent communication models assume reliable point-to-point messages passing with negligible communication costs. In particular, KQML assumes point-to-point message passing, possibly with the aid of facilitator agents [10]. Nonetheless, KQML performatives could be used for the content portions of our communication scheme. KQML does not address the problems raised by having a single, low-bandwidth communication channel.
When communication is reliable and the cost of communication relative to other actions is small, agents have the luxury of using reliable, multi-step negotiation protocols. For example, in Cohen's convoy example [7], the communication time required to form and maintain a convoy of vehicles is insignificant compared to the time it takes the convoy to drive to its destination. Similarly, message passing among distributed information agents is typically very quick compared to the searches and services that they are performing. Thus, it makes sense for agents to initiate and confirm their coalition while guaranteeing that they will inform each other if they have trouble fulfilling their part of the joint action.
With only a single team present, a situation similar to the one considered here is examined in [21]. In that case, like in the soccer server, messages sent are only heard by agents within a circular region of the sender. Communication is used for cooperation and for knowledge sharing. Like in the examples presented in the soccer domain, agents attempt to update each other on their own internal states when communicating. However, the exploration task considered there is much simpler than soccer, particularly in that there are no opponents using the same communication channel and in that the nature of the task allows for simpler, less urgent communication.
Although communication in the presence of hostile agents is well studied in military contexts from the standpoint of encryption, the problem considered here is not the same. While any encryption scheme could be used for the message content, the work presented here assumes that the adversaries cannot decode the message. Even so, they can disrupt communication by mimicking past messages textually: presumably past message have some meaning to the team that uttered them. Our method of message coding based on a globally accessible clock circumvents this latter problem.
Even when communication time is insignificant compared to action execution, such as in a helicopter fighting domain, it can be risky for agents to rely on communication. As pointed out in [38], if the teammate with whom an agent is communicating gets shot down, the agent could be incapacitated if it requires a response from the teammate. This work also considers the cost of communication in terms of risking opponent eavesdropping and the benefits of communication in terms of shifting roles among team members. However, the problems raised by a single communication channel and the possibility of active interference are not considered, nor are the challenges raised when communication conflicts with real-time action.
Another approach that recognizes the danger of basing behaviors upon multi-step communication protocols is ALLIANCE [26]. Since a primary goal of this work is fault-tolerance, only broadcast communications are used. Agents inform each other of what they are currently doing, but never ask for responses. In ALLIANCE, the team uses time-slice communication so that each agent periodically gets exclusive use of the single communication channel.
A possible application of our communication method is to robots using audio communication. This type of communication is inherently single-channel and low-bandwidth. An example of such a system is the Robot Entertainment Systems which uses a tonal language [11]. Agents can communicate by emitting and recognizing a range of audible pitches. In such a system, the number of bits per message would have to be lowered, but the general techniques presented in this article still apply.
Another example of such a communication environment is arbitrarily expandable systems. If agents aren't aware of what other agents exist in the environment, then all agents must use a single universally-known communication channel, at least in order to initiate communication.