'Continuous-action reinforcement learning' - COMP 396 Undergraduate Research Project Application Form

Supervisor's Name: Doina Precup

Supervisor's Email: dprecup [at] cs.mcgill.ca

Supervisor's Phone:

Supervisor's Website:

Supervisor's department: Computer Science

Course number: COMP 396 (Computer Science)

Term: Summer 2013

Project start date: May 1, 2013

Project end date: August 31, 2013

Project title: Continuous-action reinforcement learning

Project description (50-100 words suggested): Many real-world control problems use continuous actions, but the solution adopted most often in reinforcement learning computational methods is to discretize the action space. This project's goal is to compare discretization with using continuous actions, either in an actor-critic model, or by using an action-value function with dual continuous inputs. The student will research existing methods, and devise experiments on a simulated average-reward control problem to test them empirically. The student will provide a data analysis, including statistical significance of the results.

Prerequisite: 1 term completed at McGill + permission of instructor.

Grading scheme (The final report must be worth at least 50% of final grade): 100% final report.

Project status: This project is taken. The professor has no more '396' projects this term.

How students can apply: N/A; this project is filled.

Ethics, safety, and training: Supervisors are responsible for the ethics and safety compliance of undergraduate students. This project involves NEITHER animal subjects, nor human subjects, nor biohazardous substances, nor radioactive materials, nor handling chemicals, nor using lasers.