next up previous
Next: Perception Hierarchies Up: Research Directions Previous: Research Directions

Analysis and Design of Intelligent, Hybrid Control Systems

We are interested in the design of multi-agent systems that deliver high levels of mission reliability in dynamic and rapidly evolving environments. Multi-agent systems that coordinate intelligently to optimize a system-wide objective offer interesting possibilities for achieving massive mission reliability through enhanced survivability and reusability, if they are co-ordinated in a distributed control paradigm together with decentralization of information. The decentralization of control intelligence and information suggests the possibility of high tolerance to effector or sensor degradation in individual agents through rapid and variable dynamic reconfiguration of inter-agent co-ordination protocols and individual agent operating modes. We have had a great deal of experience at Berkeley on two large civilian (or dual use) test beds: the first for Automated Highway Systems (AHS) and the other for Air Traffic Management Systems (ATM) in multi-agent systems. It is this experience that convinces us that the design of massively reliable muti-agent systems is a difficult problem. It needs research into a new system-theoretic paradigm that we broadly characterize as intelligent hierarchical semi-autonomous agent control.

In a large spatially distributed theater of operations, the sensory systems of individual agents are, at best, able to obtain localized and often, highly uncertain information though mission objectives demand that the agents act quickly, decisively and in concert to optimize global objectives. It is illuminating to decompose the process that maps sensor information to control actions in two steps. The first is the mapping of information about the uncertain, partially modeled, internal and external environment of the agent into a top level control decision. This process has been variously characterized as goal oriented planning, perceptual reasoning, optimal decision making in stochastic control, reasoning in Bayesian belief networks, fuzzy decision making, and pattern recognition in neural networks. We shall collectively refer to these as soft computing techniques. The second is the process that maps the top level control decision to the sequence of control and co-ordination actions that cascade through the multi-agent system, and ultimately result in the activation of various agent effectors. This second process is to addressed by research on verification and control of discrete event systems and hybrid systems, including protocol design. The design of massively reliable multi-agent systems requires research into a paradigm that provides for the seamless integration of soft and formal computing techniques to design a system that conforms to a top-level mission performance and reliability specification.


next up previous
Next: Perception Hierarchies Up: Research Directions Previous: Research Directions

S Sastry
Sun Aug 9 16:58:51 PDT 1998