300
Logics for Multiagent Systems
specification languages. The theoretical complexity of proof for many of these logics is quite daunting. Hindriks and colleagues have used Plotlun's structured operational semantics to axiomatize their 3APL language (Hindriks et al., 1998, 1999). With respect to model-checking approaches, the main problem, as we indicated above, is again the issue of ungrounded semantics for agent specification languages. If we cannot take an arbitrary program and say, for t h s program, what its beliefs, desires, and intentions are, then it is not clear how we might verify that this program satisfied a specification expressed in terms of such constructs. Formalisms for reasoning about agents have come a long way since Hintikka's pioneering work on logics of knowledge and belief (Hintikka, 1962).Within AI, perhaps the main emphasis of subsequent work has been on attempting to develop formalisms that capture the relationship between the various elements that comprise an agent's cognitive state; the paradigm example of this work is the wellknown theory of intention developed by Cohen and Levesque (1990a).Despite the very real progress that has been made, there still remain many fairly fundamental problems and issues outstanding. On a technical level, we can identify a number of issues that remain open. First, the problems associated with possible-worlds semantics (notably, logical omniscience) cannot be regarded as solved. As we observed above, possible worlds remain the semantics of choice for many researchers, and yet they do not in general represent a realistic model of agents with limited resources - and of course all real agents are resource-bounded. One solution is to ground possible-worlds semantics, giving them a precise interpretation in terms of the world. T h s was the approach taken in Rosenschein and Kaelbling's situated automata paradigm, and can be very successful. However, it is not clear how such a grounding could be given to pro-attitudes such as desires or intentions (although some attempts have been made (Singh, 1990a; Wooldridge, 1992; Werner, 1990)).There is obviously much work remaining to be done on formalisms for knowledge and belief, in particular in the area of modelling resource-bounded reasoners. With respect to logics that combine different attitudes, perhaps the most important problems still outstanding relate to intention. In particular, the relationship between intention and action has not been formally represented in a satisfactory way. The problem seems to be that having an intention to act makes it more likely that an agent will act, but does not generally guarantee it. While it seems straightforward to build systems that appear to have intentions (Wooldridge, 1995), it seems much harder to capture this relationship formally. Other problems that have not yet really been addressed in the literature include the management of multiple, possibly conflicting intentions, and the formation, scheduling, and reconsideration of intentions. The question of exactly which combination of attitudes is required to characterize an agent is also the subject of some debate. As we observed above, a currently popular approach is to use a combination of beliefs, desires, and intentions