| Popis: |
The rapid growth in the deployment of autonomous systems across various sectors has generated considerable interest in how these systems can operate reliably in large, stochastic, and unstructured environments. Despite recent advances in artificial intelligence and machine learning, it is challenging to assure that autonomous systems will operate reliably in the open world. One of the causes of unreliable behavior is the impreciseness of the model used for decision-making. Due to the practical challenges in data collection and precise model specification, autonomous systems often operate based on models that do not represent all the details in the environment. Even if the system has access to a comprehensive decision-making model that accounts for all the details in the environment and all possible scenarios the agent may encounter, it may be intractable to solve this complex model optimally. Consequently, this complex, high fidelity model may be simplified to accelerate planning, introducing imprecision. Reasoning with such imprecise models affects the reliability of autonomous systems. A system's actions may sometimes produce unexpected, undesirable consequences, which are often identified after deployment. How can we design autonomous systems that can operate reliably in the presence of uncertainty and model imprecision? This dissertation presents solutions to address three classes of model imprecision in a Markov decision process, along with an analysis of the conditions under which bounded-performance can be guaranteed. First, an adaptive outcome selection approach is introduced to devise risk-aware reduced models of the environment that efficiently balance the trade-off between model simplicity and fidelity, to accelerate planning in resource-constrained settings. Second, a framework that extends stochastic shortest path framework to problems with imperfect information about the goal state during planning is introduced, along with two solution approaches to solve this problem. Finally, two complementary solution approaches are presented to minimize the negative side effects of agent actions. The techniques presented in this dissertation enable an autonomous system to detect and mitigate undesirable behavior, without redesigning the model entirely. |
