Goal Context | Body
5.5 Alternate reasoning methods
5.5.2 Practical agent reasoning
5.5.2.1 Hybrid agents
In the last chapter the set of required capabilities for agent systems was specified as: autonomy, reactivity, deliberation and sociability. The processing for these capabilities requires differing resources that are not always complementary. The limitation with deliberative agents is that they are not able to respond quickly to changes or unexpected events in their environment. The limitation with reactive agents is that they are not really capable of pursing a range of goals over a long term. One solution to the differing requirements is to allow different subsystems, or layers, to process the deliberative and reactive requirements separately and then combine results to provide the overall agent behaviour (Müller 1991).
This hybrid arrangement of processing layers allows the agent to produce timely responses to changes in the environment while pursuing longer- term goals. Hybrid agents such as INTERRAP (Müller 1991) and Touring
Figure 5.5 Horizontal (a) and vertical (b) information flows in layered agent architecture (Courtesy of Müller 1991)
Reactive layer Deliberative layer Model layer Perceptual input Action output (a) Reactive layer Deliberative layer Model layer Perceptual input Action output (b)
79
Machines (Ferguson 1992) are typically constructed with a reactive rapid responding layer, a goal seeking deliberative layer and a third domain specific modelling layer. The major difference between the types of agent is how the layers interact. In INTERRAP the layers are arranged vertically in a hierarchy. All sensory input and action output to the environment is through the reactive layer. If an input requires more processing it can be passed up to the deliberative layer and so on to the model layer, see figure 5.5 (b). If a layer is able to process an item of information the result is passed down the hierarchy where it may affect the operation of a lower layer or produce an action via the reactive layer. In Touring Machines the layers are arranged horizontally. Each layer has sensory input and action output to the environment, see figure 5.5 (a). Information in the agent is processed in parallel by each layer; because of this it is possible for layers to produce contradictory actions so each layer contains a mediation function to inhibit, or be inhibited by, other layers giving control to one layer only at any particular time (Ferguson 1992). The horizontal reasoning, Touring Machines architecture, makes use of suppression and inhibition mechanisms similar to that used in the subsumption architecture to determine which layer controls the agent’s behaviour.
Cooperative Planning Layer
Local Planning Layer
Behavior-Based Layer Social Model
Mental Model
World Model
Sensors Communications Actions
Figure 5.6 The Interrap agent architecture (Courtesy of Müller 1991)
SG DE
SG DE
80
The INTERRAP architecture consists of three vertically layered processing areas that each process perceptions from its environment at a different level of abstraction, see figure 5.6. Each layer consists of two processes called SG, for recognising situations and setting goals and the DE process for making decisions and overseeing plan execution. The lowest layer, called the behaviour based layer (BBL), deals with supervising reactive responses to changes in the environment. The middle layer, called the local planning layer (LPL), implements a planner to generate plans required to achieve the proactive goals of the agent. The highest layer, the cooperative planning layer (CPL), governs social interactions with other agents.
Another example of a layered hybrid system is the Prodigy/RAPS architecture developed by Veloso and Rizzo (1998). This consists of two separate reasoning layers. The upper layer is Prodigy, which is a deliberative reasoning system, although it is not clear from the authors whether or not the Prodigy planner includes the Analogy CBR engine for this architecture. The lower layer is based on James Firby’s Reaction Action Package system (RAPS), a rule processor, which executes planning goals that are specified as knowledge structures similar to the reactive plans of a BDI architecture. Plans generated by Prodigy are translated into RAPS operators, as the two systems do not share a common syntax, for execution where RAPS controls the pursuit of deliberative and reactive goals without intervention from Prodigy. Another hybrid architecture, called CBR-BDI, combines a BDI planner with a CBR to address some of the limitations of BDI such as the absence of a learning mechanism, the need to recompile the agent knowledge base to add new plans and the efficiency of some implementations (Bajo & Corchado 2005). The architecture is not layered but rather implements the BDI reasoning within the CBR by mapping the BDI knowledge structures onto the cases in the knowledgebase. In a CBR-BDI a case represents the set of beliefs, an intention and a desire, which cause the resolution of a problem (Corchado & Pellicer 2005). The mapping between cases and BDI plans are for the problem component of a case to represent the beliefs, the solution component is equivalent to the
81
intentions and the result represents the desires. Reasoning in the CBR- BDI is performed in the four REs process cycle of the CBR engine. It is not clear from the authors how efficiently reactive processing is supported in the architecture compared to other BDI systems, however the agent is able to reason, communicate and learn.