APT Agents: Agents that are Adaptive, Predictable, and Timely

Agents, such as robots or softbots, are becoming an increasingly prevalent paradigm. Many software systems of the future will be multi-agent systems. The agents paradigm offers numerous advantages, including increased flexibility, fault-tolerance and security. Nevertheless, it can also introduce problems due to the complexity and sophistication of multi-agent systems. The solution that we've developed is APT agents, i.e., agents that are simultaneously adaptive, predictable and timely.

Project Description

This project has originated the notion of "safe" machine learning systems, i.e., systems that have the flexibility to adapt to unforeseen circumstances by learning, but are guaranteed to remain within critical behavioral constraints. The primary focus has been on developing time-efficient methods for ensuring the safety of learning -- because it is assumed that the systems will need to adapt quickly while online.

Our results consist, in part, of a priori guarantees that certain machine learning methods preserve important classes of properties (constraints). This implies that systems can use these learning methods with no run-time cost for re-verifying that the properties still hold subsequent to learning (assuming they held prior to learning). For learning methods that are not a priori guaranteed to preserve properties, we have developed time-efficient incremental re-verification algorithms to determine online whether special cases of these methods are safe.

Currently, the method for verifying the adaptive systems is model checking. In the future, we plan to explore probabilistic model checking and monitoring/checking of adaptive systems.

Safe learning systems will be crucial in a wide range of applications. We are particularly interested in ensuring the safety of adaptive multi-agent systems, e.g.: (1) a group of cooperating planetary rovers that can quickly learn to adapt to unforeseen conditions but need to remain within mission constraints, (2) anti-viruses that can evolve to more effectively combat viruses without also evolving virus-like behavior, and (3) intrusion detection systems with countermeasures that improve but don't deny service to friendly users.

Recent Publications

For more information, please contact Diana Gordon-Spears.
Last modified: 04/17/01