Kurt D. Krebsbach

Self-adaptive software for hard real-time environments

Researchers in the Automated Reasoning group at the Honeywell Technology Center and at the University of Michigan are developing adaptive intelligent software for high-risk situations. We are building a system called Self-Adaptive CIRCA (based on our cooperative intelligent real-time control architecture model) that combines the assurance of hard real-time systems with the self-modeling, self-monitoring, and self-modifying capabilities of self-adaptive software. The article describes elements of the system that are working now, as well as new components that we are still in the process of designing and building.

An empirical study of sensing and defaulting in planning

Traditional approaches to task planning assume that the planner has access to all of the world information needed to develop a complete, correct plan which can then be executed in its entirety by an agent. Since this assumption does not typically hold in realistic domains, we have implemented a planner which can plan to perform sensor operations to allow an agent to gather the information necessary to complete planning and achieve its goals in the face of missing or uncertain environmental information. Naturally this approach requires some execution to be interleaved with the planning process. In this paper we present the results of a systematic experimental study of this planners performance under various conditions. The chief difficulty arises when the agent performs actions which interfere with or, in the worst case, preclude the possibility of the achievement of its later goals. We have found that by making intelligent decisions about goal ordering, what to sense, and when to sense it, the planner can significantly reduce the risk of committing to premature action. We have studied the problem both from the perspective of reversible and irreversible actions.

Building Coordinated Real-Time Control Plans

We are interested in developing multi-agent systems that can provide real-time performance guarantees for critical missions that require cooperation. In particular, we are developing methods for teams of CIRCA agents to build coordinated plans that include explicit runtime communications to support distributed real-time reactivity to the environment. These teams can build plans in which different agents use their unique capabilities to guarantee that the team will respond in a coordinated fashion to mission-critical events. By reasoning explicitly about different agent roles, the agents can identify what communications must be exchanged in different situations. And, by reasoning explicitly about domain deadlines and communication time, the agents can build reactive plans that provide end-to-end performance guarantees spanning multi-agent teams.

Deliberation scheduling using GSMDPs in stochastic asynchronous domains

We propose a new decision-theoretic approach for solving execution-time deliberation scheduling problems using recent advances in Generalized Semi-Markov Decision Processes (GSMDPs). In particular, we use GSMDPs to more accurately model domains in which planning and execution occur concurrently, plan improvement actions have uncertain effects and duration, and events (such as threats) occur asynchronously and stochastically. In this way, agents develop a continuous-time deliberation policy offline which can then be consulted to dynamically select deliberation-level and domain-level actions at plan execution-time. We demonstrate a significant improvement in expressibility over previous discrete-time approximate models in which mission phase duration was fixed, failure events were synchronized with phase transitions, and planning time was discretized into constant-sized planning quanta.