Once again, there are many possible ways in the current scenario to enhance MAS with ML techniques. Within this heterogeneous communicating multiagent scenario there is a clear need to pre-define a language and communication protocol for use by the agents. However, an interesting alternative would be to allow the agents to learn for themselves what to communicate and how to interpret it. For example, an agent might be given a small language of utterances and a small set of meanings, but no mapping between the two. Agents would then have to learn both what to say and how to interpret what they hear. A possible result would be more efficient communications: they would need to be understandable only by the agents rather than by both agents and humans.
When considering communications as speech acts, agents could be allowed to learn the effects of speech on the global dynamics of the system. In domains with low bandwidth or large time delays associated with communication, the utility of communicating at a given moment might be learned. In addition, if allowed to learn to communicate, agents are more likely to avoid being reliably conned by untruthfulness in communication: when another agent says something that turns out not to be true, it will not be believed so readily in the future.
Finally, commitment--the act of taking on another agent's goals--has both benefits and disadvantages. System builders may want to allow their agents to learn when to commit to others. The learning opportunities in this scenario are summarized in Table 7.