With the addition of communication as a capability available to agents, it is possible to consider this capability as an ``action'' no different from any other. When an agent transmits information to another agent, it has an effect just like any other action would have. Thus within a planning framework, one can define preconditions and effects for communicative acts. When combined with a model of other agents, the effect of a communication act might be to alter an agent's belief about the state of another agent or agents. The theory of communication as action is called speech acts [14, 51].
Mataric adds a learning dimension to the idea of speech acts. Starting with the foraging behavior mentioned above [52], the agents can then learn to choose among a set of social behaviors that include broadcasting and listening [53]. Q-learning is extended so that reinforcement can be received for direct rewards or for rewards to other agents.
When using communication as a planning action, the possibility arises of communicating misinformation in order to satisfy a particular goal. For instance, an agent may want another agent to believe that something is true. Rather than actually making it true, the agent might just say that it is true. For example, Sandholm and Lesser analyze a framework in which agents are allowed to ``decommit'' from agreements with other agents by paying a penalty to these other agents [74]. They consider the case in which an agent might not be truthful in its decommitment, hoping that the other agent will decommit first. In such situations, agents must also consider what communications to believe /citeRosenschein94:Rules.