Genetic algorithms for action set selection across domains: a demonstration

Action set selection in Markov Decision Processes (MDPs) is an area of research that has received little attention. On the other hand, the set of actions available to an MDP agent can have a significant impact on the ability of the agent to gain optimal rewards. Last year at GECCO'05, the first automated action set selection tool powered by genetic algorithms was presented. The demonstration of its capabilities, though intriguing, was limited to a single domain. In this paper, we apply the tool to a more challenging problem of oil sand image interpretation. In the new experiments, genetic algorithms evolved a compact high-performance set of image processing operators, decreasing interpretation time by 98% while improving image interpretation accuracy by 55%. These results exceed the original performance and suggest certain cross-domain portability of the approach. Categories and Subject Descriptors I.4.8 [Computing Methodologies]: Artificial Intelligence--Learning General Terms P...

• Type: Conference
• Year: 2006
• Where: GECCO
• Authors: Greg Lee, Vadim Bulitko
• Keywords: Action Set Selection, GECCO 2006, Genetic Algorithms, Image Interpretation, Optimization