Browsing by Subject "Action selection"
Now showing 1 - 2 of 2
Results Per Page
Sort Options
Item Action selection and coordination of autonomous agents for UAV surveillance(2011-12) Han, David Ching-Wey; Barber, K. Suzanne; Arapostathis, Aristotle; Aziz, Adnan; Lifschitz, Vladimir; Stone, PeterAgents, by definition, (1) are situated in an environment upon which their actions affect changes and (2) have some level of autonomy from the control of humans or other agents. Being situated requires that the agent have a mechanism for sensing the environment as well as actuators for changing the environment. Autonomy implies that each agent has the freedom to make their own decisions. Rational agents are those agents that decide to execute actions that are in their “best interests” according to their desires, using a model of those desires on which they make those decisions. Action selection is complicated due to uncertainty when operating in a dynamic environment or where other actors (agents) can also influence the environment. This dissertation presents an action selection framework and algorithms that are (1) rational with respect to multiple desires and (2) responsive with respect to changing desires. Agents can use the concept of commitments, and the subsequent communication of those commitments, to coordinate their actions and reduce their uncertainty. Coordination is layered on top of this framework by describing and analyzing how commitments affect the agents’ desires in their action selection models. This research uses the domain of UAV surveillance to experimentally explore the balance between under-commitment and over-commitment. Where previous approaches concentrate on the semantics of commitment, this research concentrates on the pragmatics of commitment, describing how to use utility calculations to enable an agent to decide when making a commitment is in its best interests.Item Action selection in modular reinforcement learning(2014-08) Zhang, Ruohan; Ballard, Dana H. (Dana Harry), 1946-Modular reinforcement learning is an approach to resolve the curse of dimensionality problem in traditional reinforcement learning. We design and implement a modular reinforcement learning algorithm, which is based on three major components: Markov decision process decomposition, module training, and global action selection. We define and formalize module class and module instance concepts in decomposition step. Under our framework of decomposition, we train each modules efficiently using SARSA($\lambda$) algorithm. Then we design, implement, test, and compare three action selection algorithms based on different heuristics: Module Combination, Module Selection, and Module Voting. For last two algorithms, we propose a method to calculate module weights efficiently, by using standard deviation of Q-values of each module. We show that Module Combination and Module Voting algorithms produce satisfactory performance in our test domain.