Reasoning and Learning Lab, McGill University, February 2014-February 2015.
Extracting features from Markov decision processes (MDPs) is an important technique for improving the performance of planning algorithms like value iteration. The hope is that if one can extract pertinent features one can plan actions based on these features rather than on detailed knowledge of the state space. In this project, we are using techniques based on probabilistic bisimulation to identify useful features. We want to define a suitable notion of a distance from a state to a feature and use it to identify features. The project involves both theoretical contributions and experimental work.
- Representation Discovery for MDPs Using Bisimulation Metrics. The Twenty-Ninth AAAI Conference on Artificial Intelligence (AAAI). 2015. [Slides] [Poster]
- Bisimulation Metric Computation for Markov Decision Processes. The 10th Undergraduate Computer Science Research Symposium (UCORE’14), Montreal, QC, Canada, August 29, 2014. [Abstract] [Slides]
- Bisimulation Metric Computation for Markov Decision Processes. Reasoning and Learning Lab, McGill University, Montreal, QC, Canada, August 20, 2014. [Slides]
- On Using Earth Mover’s Distances for Bisimulation Metrics. Reasoning and Learning Lab, McGill University, Montreal, QC, Canada, June 25, 2014. [Slides]