We believe that qualitative leaps in scope and performance will emerge from addressing the basic problems together. Complexity and spatial extent are addressed by system decomposition based on hierarchical, hybrid, and multi-agent designs, using multiple levels of abstraction for sensory and control functions. Structural and parametric learning methods adapt the system to initially unknown environments, while generalized estimation methods, uncertainty management, and robust control techniques cope with the residual uncertainty inherent in stochastic, partially observable environments. Real-time decision-making is achieved by parallelism, reflexive control, compilation, and anytime approximation algorithms. Above and beyond the development of these specific methods, we see the possibility for a reunification of control, AI, and computational neuroscience into a theoretical and technological continuum, with enormous benefits for the science and engineering of intelligent systems.

1. Control, AI, and Computational Neuroscience

1. Control Theory made significant strides in developing algorithms for control systems design using optimal, adaptive, and linear and nonlinear control techniques. Comparatively less was done about run-time issues, choice of sensors and actuators (which is done by the experienced designer), and the interaction between planning and control. Key areas of future research are in hybrid systems, distributed perception and control using hierarchical / heterarchical architectures, and design techniques for the integration of high level reasoning with real-time algorithms.
   
   
2. Artificial Intelligence has made progress in perception, learning, representing knowledge, inferencing, and planning, to the point where the deployment of integrated intelligent agents for real environments can be contemplated. Unlike control systems, AI systems are often designed to incorporate and process explicit domain models. (Particularly important are domain models incorporating uncertainty, such as neural networks, Bayesian belief networks, and fuzzy logic.) From these models, plans of action are constructed for specific tasks and situations. Plan execution and monitoring is the counterpart of ``real-time control.'' Since inferencing can be computationally intractable, recent research has included the design of reactive systems (roughly equivalent to ``control laws'') and compilation (the counterpart of architectural design).
   
   
3. Computational Neuroscience and Cognitive Sciences focused largely on the study of biological systems. Sensorimotor control studies adapted intelligent control techniques to the study of locomotory systems for insects, oculomotor systems for humans, etc. Reasoning, inference, and problem solving are studied in an information processing / cognitive science framework which is close to the artificial intelligence paradigm, whereas perception systems have been studied from both points of view. Recently, the development of mathematical models for the storage and processing of complex perceptual data using neural networks has come to the forefront.

2. New Challenges: Intelligent Multi-Agent Systems

Although a Central Control paradigm no longer applies here, control engineers have with great success used its theories and its design and analysis tools to build and operate these distributed control systems. There are two reasons why the paradigm succeeded in practice, even when it failed in principle. First, in each case the complexity and scale of the material process grew incrementally and relatively slowly. Each new increment to the process was controlled using the paradigm, and adjustments were slowly made after extensive (but by no means exhaustive) testing to ensure that the new controller worked in relative harmony with the existing controllers. Second, the processes were operated with a considerable degree of ``slack.'' That is, the process was operated well within its performance limits to permit errors in the extrapolation of test results to untested situations and to tolerate a small degree of disharmony among the controllers. However, in each system mentioned above, there were occasions when the material process was stressed to its limits and the disharmony became intolerable, leading to a spectacular loss of efficiency. For example, most air travelers have experienced delays as congestion in one part of the country is transmitted by the control system to other parts. The distributed control system of the interconnected power grid has sometimes failed to respond correctly and caused a small fault in one part of a grid to escalate into a system-wide blackout.

We are now attempting to build control systems for processes that are vastly more complex or that are to be operated much closer to their performance limits in order to achieve much greater efficiency of resource use. The attempt to use the central control paradigm cannot meet this challenge: the material process is already given and it is not practicable to approach its complexity in an incremental fashion as before. Moreover, the communication and computation costs in the central control paradigm would be prohibitive, especially if we insist that the control algorithms be fault-tolerant. What we need to meet the challenge of control design for a complex, high performance material process, is, we believe, a new paradigm for distributed control. It must distribute the control functions in a way that avoids the high communication and computation costs of central control, at the same time that it limits complexity. The distributed control must, nevertheless, permit centralized authority over those aspects of the material process that are necessary to achieve the high performance goals. We believe that such a challenge can be met by organizing the distributed control functions in a hierarchical architecture that makes those functions relatively autonomous (which permits using all the tools of central control), while introducing enough coordination and supervision to ensure the harmony of the distributed controllers necessary for high performance.

This is an ambitious project, but we feel that the time is ripe for a synthesis of control theory, artificial intelligence, and biological information processing into a new paradigm for the design of intelligent systems. It would be foolhardy to start from a tabula rasa. On the contrary, we will use all of the accumulated knowledge and experience of the central control paradigm: an exciting orchestration of ideas from geometric control theory, decision theory, dynamical systems, discrete event and hybrid systems, perceptual systems, soft computing,, formal verification, machine learning and adaptive control, game theory, logical and probabilistic reasoning, real-time decision making, and cognitive modelling.

In addition synthesizing these diverse approaches and experiences into a unified paradigm, we will confirm or validate this new paradigm by using it for controlling our test processes. Thus, our program follows the classical pattern of scientific progress: the first phase of ``induction'' or the integration of approaches and experiences that go beyond the current practice into a new paradigm which subsumes the current one; and the second phase of ``deduction'' or the application of the new paradigm to concrete situations to test its validity. We will be guided in our choice of problems by a number of detailed case studies of large, complex systems with multiple agents arising in intelligent vehicle highway systems, air traffic management systems, and intelligent telemedicine with both experimental and simulation testbeds. An important new feature of our approach is that we will work on both fully autonomous intelligent systems and intelligence augmentation of human centered systems. The latter systems raise special cognitive issues relating to interaction with humans. In this context, intelligent software agents can provide significant performance enhancement for human operators of cars, aircraft, and other complex systems.