We view the key research areas to be the following:The following subsections outline each of these five areas.
- Design of hierachical control architectures for multi-agent systems that share a single environment
- Perception systems:
- hierarchical aggregation
- wide area surveillance
- low-level perception
- Frameworks for representing and reasoning with uncertainty
- Incorporation of learning and adaptation in the agents
- Soft computing approaches to intelligence augmentation for human centered systems
1. Analysis and Design of Multi-agent Hybrid Control Systems
One of the main incentives to move into the area of multi-agent large scale systems is economic. Preliminary studies indicate that automation can improve coordination in air traffic management systems, highway systems, chemical process control, power generation and distribution, etc. This in turn leads to performance improvement in terms of fuel consumption, safety, efficiency, and environmental impact. Further, in Army applications, such as intelligent battlefield telemedicine and battlefield management, it is critical to organize the flow of perceptual information and sensory actions so as to increase battlefield awareness without sensory overload. To deal with complex systems, engineers use a combination of continuous and discrete controllers. Continuous controllers are used primarily because interaction with the physical plant, through sensors and actuators, is essentially analog, and continuous models and design techniques have been developed, used, and validated extensively. An equally compelling case exists for discrete controllers: since discrete abstractions make it easier to manage system complexity, discrete models are easier to manipulate, and discrete abstractions more naturally accommodate linguistic and qualitative information in the controller design. We will use the term ``hybrid systems'' to describe systems that incorporate both continuous and discrete dynamics. Our goal is to derive a useful, general approach to the design of hybrid controllers for complex systems.Multi-Agent Scarce Resource Systems
An important class of systems that are well suited for hybrid control are multi-agent, scarce resource systems. Their common characteristic is that many agents are trying to make use of a common, congestible resource. For example, in highway systems, the vehicles are agents competing for scarce highway space-time resources, while in air traffic management systems the aircraft compete for air space and runway space.To achieve the common optimum we should ideally have a centralized control scheme that computes the global optimum and commands the agents accordingly. A solution like this may be undesirable, however, for several reasons: 1) it is likely to be very computationally intensive, as a large centralized computer is needed to make all the decisions; 2) it may be less reliable, as the consequences may be catastrophic if the centralized controller is disabled; 3) the information that needs to be exchanged may be too expensive; and 4) the number of agents may be large and/or dynamically changing.
If the performance degradation of a completely decentralized solution is unacceptable and a completely centralized solution is prohibitively complex or expensive, a compromise will have to be found. Such a compromise will feature semi-autonomous agent operation. In this case, each agent is trying to optimize its own usage of the resource and coordinates with ``neighboring'' agents in case there is a conflict of objectives. It should be noted that semiautonomous agent control is naturally suited for hybrid designs. At the continuous level, each agent chooses its own optimal strategy, while discrete coordination is used to resolve conflicts. Thus, the class of hybrid systems that we will be most interested in are multi-agent systems, where the hybrid dynamics arise from the interaction between continuous single agent ``optimal'' strategies and discrete conflict resolution or coordination protocols.
Architectures for Multi-Agent Hybrid Control Systems
There is a continuum of design choices for system decomposition, ranging from strict hierarchical control to a fully distributed, multi-agent system. Furthermore, different choices may be appropriate at different levels of abstraction, ranging from the (typically continuous-domain) low-level control systems concerned with safety and smooth execution to the (typically symbolic/discrete) strategic levels concerned with optimization and planning for high-level goals. We will investigate theoretical and design issues involved in the choice of system architecture, and methods for interfacing elements of the resulting hybrid system. We will study the choice of hierarchies and heterarchies required for individual systems, based on a detailed consideration of Intelligent Vehicle Highway Systems, Air Traffic Management Systems, and Intelligent Battlefield Telemedicine. We will investigate the extent to which we can organize the control of complex systems to achieve emergent optimal behavior of the collective system for the usage of a scarce resource by many independent agents.Top
2. Perceptual Data Abstraction
A sensorimotor task is characterized by an upward hierarchy of abstractions of perceptual data and a downward hierarchy of elaborations for motor control. Thus, a key requirement of hierarchical intelligent systems is their ability to abstract multi-sensory perceptions with labels for the planning layers of the hierarchy, and a capability to expand high-level instructions into detailed actions for all the agents. Our approach to perceptual data abstraction involves the use of both local and wide area sensing and data fusion and aggregation at different levels of granularity. Basic research on methodologies for wide area surveillance and low level vision are combined with semantic information about the data being encoded to infer behaviors of the complex multi-agent system. We will produce high-speed Cellular Neural Network implementations for low level vision algorithms, especially on our IVHS test bed.We will also use neural network models for data compression and classification of complex data patterns. A form of chaotic neural network with an array of coupled oscillators motivated by a biological model of olfactory and auditory data storage will prove valuable in this context.
Top
3. Representation and Reasoning with Uncertainty
Representation and manipulation of information using neural network models, Bayesian belief networks, or fuzzy linguistic models are needed for decision making with noisy inputs and uncertain models. We will investigate extensions in the expressive power of such formalisms needed to model complex, dynamic environments at multiple levels of abstraction. Fuzzy logic and nonmonotonic reasoning provide a framework for reasoning linguistically about conflict resolution among multiple agents. Dynamic belief networks provide a decomposed representation for dynamic, stochastic environments, and handle hybrid models. We will extend this representation to incorporate first-order expressive power (first-order probabilistic logic, knowledge-based model construction), which enables a single knowledge base to represent an entire class of specific dynamical systems. Algorithms that dynamically construct and adapt the model as circumstances change will allow the system to focus attention on the most relevant aspects of the situation. Use of hierarchical abstractions and least-commitment plan representations will provide for exponential complexity reduction in decision-making under uncertainty.Top
4. Adaptation and Learning
Learning must take place at all levels of the system, since we cannot assume that the environment and correct system structure are known at the outset. This includes learning the environment model from sensory inputs, learning value information from normative inputs, and direct learning of control laws in supervised and unsupervised settings. In addition, the hierarchical/multi-agent structure of the system must itself be adaptable. Work on the task of learning environment models from observation, particularly partially observable environments with noisy sensors, will be closely integrated with work on perception and representation. For example, we are interested in learning predictive models of human drivers directly from video recordings. Work on learning utility information and control laws from reinforcement signals will be closely integrated with work on representation and control. The use of expressively powerful first-order probabilistic models of dynamical systems will require new learning methods for building structured models and adjusting their parameters; the learning process can then be amortized over the large number of situations in which the same structural model applies, and the system will be robust with respect to parametric changes in the environment.Learning-based verification provides a link between machine learning and formal verification of hybrid systems. Formal techniques from computational learning theory that bound the error in the learned environment model or control law, together with a formal analysis of decision quality and system failure rates when the environment model is only approximately correct, yield guarantees on the performance of control systems designed to operate in unknown environments.
Top
5. Intelligence Augmentation of Human Centered Systems
We propose to extend methods for augmentation of human centered systems well beyond current low-level methods such as night vision and cruise control. This requires interfacing with higher-level human cognitive functions-providing information tailored to the human perceptual system, accepting commands in imprecise and linguistic forms, and adapting to user preferences. For these tasks, methods from Soft Computing (fuzzy logic, neural networks, genetic algorithms, and probabilistic reasoning) seem most appropriate. Key to success will be techniques for combining these methods into hybrid systems that be able to represent and adapt the required mappings from human perceptual and preference space into the underlying system parameter space.
[Top|Home]