β Scribed by Barbara Hayes-Roth
No coin nor oath required. For personal study only.
How should an artificially intelligent agent decide which action to perform at each point in time?
In my 1985 paper, "A blackboard architecture for control" [7], I claimed that the control problem is fundamental to all cognitive processes and intelligent systems. At each point in time, many actions may be logically possible, given the current state of the task environment. An intelligent agent must choose among them, either implicitly or explicitly. For a given task, the agent's solution to the control problem determines what actions it performs and, therefore, what goals it achieves, what resources it uses, and what side-effects it produces. More generally, the agent's approach to control determines its potential range of behavior, its run-time flexibility, and its coherence and comprehensibility to observers.
Therefore I argued that, like human beings--and unlike conventional computer programs--an AI agent must have capabilities for intelligent control. It must use knowledge and reasoning skills to construct and modify explicit control plans for its own actions at run time. It should notice what actions are possible at each point in time and use its control plans to choose among them. Its control plans should be specific enough to select sequences of actions that achieve goals, yet general and flexible enough to accommodate unanticipated demands and opportunities for action. The agent should modify its control plans as warranted by events in the task environment.
To realize intelligent control in an AI agent, I proposed the blackboard control architecture shown schematically in Fig. 1.
π SIMILAR VOLUMES
A neural network controller is developed to learn the inverse dynamics of unknown dynamic systems and to serve as afeedforward controller. Arttjicial neuralnetworks, which consist of a set oj'processing units with interconnections between them, are used to get the desired output. The interconnection