In general, agents can be neither totally deliberative nor totally reactive. If they are only reactive, they cannot reason about their actions and will not be able to achieve any sophisticated behavior; if they are just deliberative they may never be able to act in time.
It is now generally recognized, therefore, that if agent systems are to survive in real and complex environments they need to be reactive in order to respond to environmental changes with sufficient speed, and be deliberative in order to achieve complex goals without deleteriously affecting longer-term options [27]. If environments change rapidly or unexpectedly, agents may need to act in a reactive manner, whereas more stable environments may allow agents time to deliberate on the best course of action. Architectures containing both deliberative and reactive components are hybrid architectures, of which two key examples, TouringMachines [4] and INTERRRAP [28], are discussed next.
2.4.1 TouringMachines
Ferguson proposed TouringMachines as an architecture for controlling and coordinating the actions of autonomous agents situated in dynamic multiagent worlds. He argued that agents operating in the real world would inevitably have to deal with several issues that had not—at that time—been considered in the vast array of architectures that exhibited quite well-defined preprogrammed behavior. Such architectures would need to address a number of possibly conflicting issues, as follows:
They need the ability to deal with unexpected events in the real (physical or electronic) world, and do so at different levels of granularity.
They need to deal with the dynamism in the environment created by the actions of other agents.
They must pay attention to environmental change.
They need to reason about temporal constraints in the knowledge that computation is necessarily resource-bounded.
They must reason about the impact the short-term actions may have on long-term goals.
In response, Ferguson proposed the TouringMachines architecture, illustrated in Figure 2.3 [27] The architecture is similar to the subsumption architecture in that it consists of a number of layers, the reactive layer, the planning layer, and the modeling layer, which continually compete to control the agent's behavior. The reactive layer responds quickly to events not explicitly programmed in the other layers, such as when a new agent or obstacle is perceived. Generating, executing, and modifying plans, such as constructing a route in order to move to a target destination, are the responsibilities of the planning layer. Finally, the modeling layer is used for building and maintaining models of entities in the environment (including the agent itself), which are used to understand the current behaviors of others and make predictions about their future behaviors.
Figure 2.3: The TouringMachines architecture. (After: [27])
Each layer models the agent's world at a different level of abstraction but, unlike the subsumption architecture, each is directly connected to both the action and perception of the agent, and any two layers can communicate with each other. Since these layers are each modeled at different abstraction levels, they each have an incomplete view of the world, encoding different strategies and tactics for responding to events. Conflicts over action-selection therefore inevitably arise, so the architecture includes a set of global context-dependent control rules to achieve coherence between these layers, which work by suppressing either the input to, or the output from, a layer.
Ferguson showed that this architecture was feasible and appropriate for agents operating in dynamic environments. Moreover, he demonstrated that it was possible to integrate expensive deliberative reasoning mechanisms with reactive behavior-based components in a resource-bounded agent. The architecture is important because it showed that it could successfully produce a wide variety of behaviors from reactive through to goal-directed, all required of a sophisticated agent embedded in a complex environment.
2.4.2 INTERRRAP
MüUller et al. [3] also argued that layered hybrid architectures are beneficial as they support the modeling of an agent's environment at different levels of abstraction, the different times taken to respond to an action (responsiveness), and the levels of knowledge and reasoning sophistication required by an agent. However, they also acknowledge that the design of individual layers is often quite personal and not guided by any general theory of agency.
In response, they proposed an architecture that combines the advantages of BDI-style architectures and those of layered ones. This architecture, called INTERRRAP, used the BDI model to guide the construction of each of the individual layers within the architecture.
The state of each layer is represented by beliefs, goals, and intentions, and each layer can receive what is known as perceived propositions. This means that each layer can perceive information from the environment, and from other layers, but this information is always in the form of simple propositions such as at (agent (A, B)). The architecture implements three basic functions at each layer, thus providing a uniform control structure for each layer. These are enumerated here:
Belief revision maps an agent's current perception and current beliefs to a set of new beliefs. For example, an agent may perceive that an object has moved, and belief revision updates the agent's current beliefs accordingly.
Situation recognition and goal activation take the current beliefs and goals and generate a new set of goals for the agent to try to achieve.
Planning and scheduling derive a new set of intentions (what the agent is committed to doing) from the current set of beliefs, goals, and intentions of the agent.
The motivation for this kind of architecture is based around the interfaces between each of the three layers, and the ways in which layers can communicate with other layers. The developers of INTERRRAP categorized Ferguson's TouringMachines architecture as horizontally layered—each layer can interact with every other layer, as well as with the perceptions and actions. In this kind of architecture, many communication paths have to be considered and any centralized control would therefore need to be very sophisticated. In response, INTERRRAP is a vertically layered architecture in which layers are configured similarly, but where communication is only possible between adjacent layers. Clearly, in this model, fewer communication channels have to be designed. Consider a system with n layers, for example. If each layer can only communicate with adjacent layers in both directions, at most 2(n−1) channels will be required. In a horizontal architecture, the first layer alone requires 2(n − 1) paths to communicate in both directions with every other layer, the next layer needs 2(n—2) paths to ensure communication with all succeeding layers, and so on. An n-layered horizontal architecture may therefore require anything up to 2n! communication paths. The disadvantage of vertically layered architectures is that for a layer to control behavior, the control must necessarily pass through all lower layers.
As indicated above, the INTERRRAP agent architecture [14] is an example of a vertically layered architecture and is shown in Figure 2.4, adapted from [3] It comprises four layers that can communicate in both directions with adjacent layers. INTERRRAP was developed to model resource-bounded autonomous agents that interact with others in dynamic environments, while implementing a pragmatic BDI architecture. The control of the agent is contained in three different hierarchical layers known as the behavior-based layer, the plan-based layer, and the cooperation-based layer. While the behavior-based layer is always carefully modeled for specific domains, the others contain more generic information pertaining to goal-directed and social behavior.
Figure 2.4: The INTERRRAP architecture. (After: [3].)
The agent's knowledge base is correspondingly split into three layers. At the lowest level, the agent's world model comprises beliefs about the environment. At the next level, the mental model contains knowledge about the agent itself, including its goals, plans, and intentions. The social model is the highest level containing information about others and the current state of joint-plans, joint-goals, and joint-intentions. These three knowledge bases are respectively accessed and updated by the behavior-based component, plan-based component, and cooperation component. The lowest level allows agents to react to unforeseen events and to quickly execute routine tasks without any explicit deliberative reasoning. The plan-based layer allows nonsocial goal-directed behavior. Plans are either hierarchical templates that call other plans or are directly executable behaviors. Finally, the social level enables the agent to interact with others by coordinating actions and forming joint-plans.
The essential operation of the agent is simple; in response to events in the environment, control spreads upwards until the appropriate level is reached and the functions described earlier are applied accordingly. The benefits of decentralizing knowledge and behavior into layers must be weighed against the cost of communication between, and coordination of, these layers; this is analogous to problems determining the number of agents required for building multiagent applications. Extending this analogy—that the design of layered architectures is analogous to designing multiagent systems by treating each layer as an autonomous agent—the contract net protocol is proposed as a useful mechanism for coordinating the concurrent execution of symbolic reasoning at different layers. We discuss the contract net in more detail later in this chapter.
2.4.3 Other Hybrid Architectures
In more recent work on building agent development tools and architectures, agents typically have both reactive and deliberative components.
For example, the Sensible Agent Testbed, a set of tools for the building, running, monitoring, and testing of complex multiagent systems [29], involves hybrid agents. The agent architecture contains a perspective modeler that maintains an explicit model of an agent's subjective view of its environment. The model includes both the behavioral (reactive) and intentional (proactive) models of itself and others. Similarly, the autonomy reasoner dynamically determines the appropriate decision-making framework for the agent. One such framework is categorized as command driven, where the agent does not function deliberatively but simply obeys orders directly from a master agent without deliberating.
In another recent example, the DECAF agent toolkit (Distributed, Environment-Centered Agent Framework), which allows a well-defined software engineering method to be used in building multiagent systems, also uses a hybrid agent architecture [30]. On one hand, the deliberative element is incorporated using a plan editor that builds a symbolic representation of a plan to achieve some objective. On the other hand, the reactive component is the executor that simply takes a set of enabled actions and produces a set of low-level commitments to very specific courses of action.
No comments:
Post a Comment