Dynamic Control Architecture (B. Hayes-Roth)

A diagram of this architecture

Philosophy

An intelligent agent must be adaptable, versitle, and exhibit coherent behavior. To be adaptive means to have the ability to respond to an event in a dynamic environment within an acceptable response time. Versitility implies that the agent has the ability to vary its responses dependant on both what it has learned and on the current environmental context. Coherency requires that all the distinct systems the agent utilizes to adapt to its environment, and all the stratigies the agent can take advantage of to achieve versitility must be integreated with a coherent overall plan of action developed by the agent.

Architecture:

The dynamic control architecture consists of a cognition system and independant perception and action systems. All subsystems operate concurrently and asynchronously and communicate through an independant but globally accessable communications interface(CI). This underlying modular structure allows for more appropriate (to the environment) response times by interacting with subsets of the environmet concurrently thereby reducing the overall complexetyeach subsystem must be able to deal with.

The input and output modules consist of limited size buffers and dynamically modifyable perceptual filters, determined by the cognition component. The limited input buffers, located in the CI, are fed information at varying rates depending on the perceptive filter. Therefore the more important the data, the more often the system will see it. However, the CI has limited buffer size and therefore if the cognitive system does not look at the buffer often enough, events may go by in the world without notice. This strategy serves to further limit the environmental complexity encountered by first focusing the agents attention (perceptual filters) and then limiting the number of environmental events which can be active at one time (limited I/O buffers). Therefore, it is important for the cognitive system to function in real time and reason about the resources available to it.

The cognition system can also be broken into two subsystems. First is the knowledge base, which contains all knowledge, including factual, procedural and reasoning stratigies for particular tasks. This system is created in an extension of the BB1 blackboard system, and all reasoning results are stored in a globally accessable uniform conceptual graph. The second is the satisfactory reasoning cycle, which itself is composed of three parts. The first is the agenda manager, responsible for identifiying and prioritizing reasoning tasks. The second is the scheduler, which interrupts the agenda manager when it is 'ready', and schedules the next best operation on the third component, the executor. Operations can have multiple possible effects, including modifications to the perceptual filters, inteded actions, new conclusions for ongoing reasoning, etc. This enables the agent to apply multiple reasoning methods to the same problem, work on multiple problems simultaniously, and to trade off the quality for a timely response.. Within the knowlege base is a control plan which is developed by the agent in accordance with its goals, and is used to focus the reasoning cycle on completing the task at hand. This includes determining which actions have priority on the agenda, when the scheduler should interrupt, and what the perceptual filters should contain. The only changes to the control plan are those made by the agent, and hence were determined necessary by the control plan and environment at that time. This provides for a global coherence of reasonining, perception, and action within the agent while still allowing a wide variety of stratigies to be applied to a specific task, including oppornutistic actions and reactions.

Agent Properties:

The dynamic control architecture is designed to interact in real time with a dynamic environment . This includes taking advantages of opportunities and meeting deadlines which present themselves. The agent thus has the ability to sacrifice quality of response for timly responses. In order to make such a tradeoff, the agent must have the ability to reason about its own resourses and allocate them approately. It also must have the meta-level reasoning abilities to select the most sufficient reasoing method for the current situation, and to decide when to stop the reasoning process and act.

Because of the combinatorial complexity of the environment, if the agent is to guarantee a maximum fixed response time, it must selectively attend to and act upon aspects of the environment. The dynamic control architecture accomplishes this through the use of perceptual filers, limited capacity I/O buffers, concurrent subsystems, and concurrent problem solving.

The system must also act coherently. It must not be distractable when it needs all of it's resourses to complete the current plan. Coherency is implemented in the dynamic control architecture directly through the use of a control plan, which guides all decision making, including those decisions pertaining to the modification of the control plan. Indirectly, coherency depends on the uniform representation of knowledge in a globally shared conceptual graph to integrate all of the architectures components.

The conceptual graph can store both declaritive and procedural forms of the same knowledge, and new knowledge can be easily added by representing it in it's delcaritive form as an extension to the conceptual graph. One limitation of this representation is that the knowlege base grows monotonically and therefore may slow the real-time responses down as it grows.

Capabilities:

The dynamic control architecture is capable of reasoning, planning, learning, sensing and behaving. It reasons by means of a reasoning cycle, which prioritizes possible actions, schedules the next action, and then executes that action. All of these steps are guided by an overall control plan which can contain explicit reasoning stratigies and can itself be explicitly reasoned about.

The dynamic control architecture incorporates dynamic control planning, which creates a control plan of temporaly organized control decisions. This plan constrains the entire cognitive system, from the reasoning cycle above to the sensing and perception system. It also determines how and what the system learns. Both factual knowledge and control knowledge can be learned. Learning mainly occurrs through the incorporation of new knowledge, and by abstracting more efficient representations of current knowledge. There is nothing in the architecture prohibiting the implementation of any specific type of learning method, but currently the system learns mainly through interaction with an expert, and by analogy of new tasks to previous stratigies for similar tasks. The percieving and behaving components of the dynamic control architecture are sepearate asynchronous modules which function concurrently with the cognitive part of the system. They are specfic to the aspects of the world they are to sence or effect. They have the ability to slectively interpret and filter incomming and outgoing data. The filters which are applied to each are dynamically modifyable by the cognitive system.

Environment and Agent Body:

The dynamic control architecture was designed to function in a complex dynamic environment where the agent is continually presented with opportunities for action, and deadlines it is required to meet. The dynamic control architecture is designed to both take advantage of these opportunities and guarantee responses (not necessarially quality responses) to environmental deadlines. The agent has imperfect knowledgeabout the environment it exists in, and therefore, in order to make such real-time responses it must, in addition to reasoning about time, reason about it's own resourses in order to determine the best feasable solution in the avaliable time. The dynamic control architecture can do this and therefore may be best classified as a sastisficing architecture for dynamic environments.

Issues:

Guardian is not always perfectly rational because it is attempting to preform in real time, and therefore will sacrifice the quality of a solution for one that meets the deadline. The hope of this behavior is that a quick action of less quality will push off the deadline far enough so that a quality solution can be found.

Theories of Intelligence:

From a meta level perspective, the dynamic control architecture can be analized both from Simon's and from Anderson's points of veiw.

From the perspective of Anderson's rational analysis, the architecture's 'Cohrence' property is an interesting signature property easily deriaveable from the environment and the goals of the system. However, as Simon would point out, the prediction that the system will follow a coherent path does not tell us how that system does the path following or what doing so imples about constraints on other functions the system can preform.

In the dynamic control architecture, coherence was implemented by the incremental construction of an abstract control plan. This control plan places limits on the computations preformed by other aspects of the system, such as focus of attention, response times, etc. Therefore, as Simon predicts we need to know the mechanism by which the signature phenomenom is implemented in order to fully predict the behaviour of the system. It is simply not enough to know that the property 'coherrence' is predicted by the environment.

Other Architectures

Back to the Title Page