Reinforcement Learning using Policy Gradients in Reproducing Kernel Hilbert Space

Add to your list(s) Download to your calendar using vCal

• Dr Guy Lever, Department of Computer Science, University College London (UCL)
• Monday 28 September 2015, 16:00-17:00
• Mechanical Engineering B05.

If you have a question about this talk, please contact Lars Kunze.

Host: Dr. Ata Kaban

I will present a system for non-parametric policy search in reproducing kernel Hilbert space for solving reinforcement learning problems. The method has many benefits over standard parametric approaches: policies can be modeled in rich function classes; there is less need to rescale the search space using, e.g., natural gradients; the policy gradient can be easily derived and estimated; the method is adaptive to the complexity of the problem. The system uses sparse-greedy approaches to function estimation both to estimate the value function, and to maintain a compact, but expressive, policy representation. I will demonstrate the method on benchmark MDPs and simulated quadrocopter navigation experiments. If time permits I will present recent extensions to second order policy search methods.

Speaker’s homepage: http://www0.cs.ucl.ac.uk/staff/G.Lever/

This talk is part of the Artificial Intelligence and Natural Computation seminars series.

This talk is included in these lists:

• Artificial Intelligence and Natural Computation seminars
• Computer Science Departmental Series
• Computer Science Distinguished Seminars
• Mechanical Engineering B05
• computer sience

Note that ex-directory lists are not shown.