The deliberative, symbolic paradigm is, at the time of writing, the dominant approach in (D)AI. This state of affairs seems likely to continue, at least for the near future. There seem to be several reasons for this. Perhaps most importantly, many symbolic AI techniques (such as rule-based systems) carry with them an associated technology and methodology that is becoming familiar to mainstream computer scientists and software engineers. Despite the well-documented problems with symbolic AI systems, this familiarity makes symbolic AI agents (such as GRATE* [Jennings, 1993b]) an attractive proposition.
With respect to reactive architectures, Ferguson suggests that:
`[T]he strength of purely non-deliberative architectures lies in their ability to exploit local patterns of activity in their current surroundings in order to generate more or less hardwired action reponses ... for a given set of stimuli. Successful operation using this method pre-supposes: (i) that the complete set of environmental stimuli required for unambiguously determining action sequences is always present and readily identifiable - in other words, that the agent's activity can be situationally determined; (ii) that the agent has no global task constraints ... which need to be reasoned about at run time; and (iii) that the agent's goal or desire system is capable of being represented implicitly in the agent's structure according to a fixed, pre-compiled ranking scheme.' [Ferguson, 1992a]
Hybrid architectures, such as the PRS, TOURINGMACHINES, INTERRAP, and COSY, are currently a very active area of work, and arguably have some advantages over both purely deliberative and purely reactive architectures. One potential difficulty with such architectures, however, is that they tend to be ad hoc in that while their structures are well-motivated from a design point of view, it is not clear that they are motivated by any deep theory. In particular, architectures such as TOURINGMACHINES, containing a number of independent activity producing layers competing with each other in real time to control the agent's activity, seem to defy attempts at formalisation. It is a matter of debate whether this need be considered a serious disadvantage, but one argument is that unless we have a good theoretical model of a particular agent or agent architecture, then we shall never really understand why it works. This is likely to make it difficult to generalise and reproduce results in varying domains.
Most introductory textbooks on AI discuss the physical symbol system hypothesis; a good recent example of such a text is [Ginsberg, 1993]. A detailed discussion of the way that this hypothesis has affected thinking in symbolic AI is provided in [Shardlow, 1990]. There are many objections to the symbolic AI paradigm, in addition to those we have outlined above. Again, introductory textbooks provide the stock criticisms and replies.
There is a wealth of material on planning and planning agents. See [Georgeff, 1987] for an overview of the state of the art in planning (as it was in 1987), [Allen et al., 1990] for a thorough collection of papers on planning, (many of the papers cited above are included), and [Wilkins, 1988] for a detailed description of SIPE, a sophisticated planning system used in a real-world application (the control of a brewery!) Another important collection of planning papers is [Georgeff and Lansky, 1986]. The book by Dean and Wellman and the book by Allen et al. contain much useful related material [Allen et al., 1991][Dean and Wellman, 1991]. There is now a regular international conference on planning; the proceedings of the first were published as [Hendler, 1992].
The collection of papers edited by Maes [Maes, 1990a] contains many interesting papers on alternatives to the symbolic AI paradigm. Kaelbling [Kaelbling, 1986] presents a clear discussion of the issues associated with developing resource-bounded rational agents, and proposes an agent architecture somewhat similar to that developed by Brooks. A proposal by Nilsson for teleo reactive programs - goal directed programs that nevertheless respond to their environment - is described in [Nilsson, 1992]. The proposal draws heavily on the situated automata paradigm; other work based on this paradigm is described in [Kiss and Reichgelt, 1992][Shoham, 1990]. Schoppers has proposed compiling plans in advance, using traditional planning techniques, in order to develop universal plans, which are essentially decision trees that can be used to efficiently determine an appropriate action in any situation [Schoppers, 1987]. Another proposal for building `reactive planners' involves the use of reactive action packages [Firby, 1987].
Other hybrid architectures are described in [Bussmann and Demazeau, 1994][Aylett and Eustace, 1994][Downs and Reichgelt, 1991].