Seminar Details
2009-12-01 (14.00) : Learning about near-optimal policies from a sample of trajectories
At Euler building
Organized by Mathematical Engineering
Speaker :
Damien Ernst (ULg)
Abstract :
In many areas of the medical, social and engineering
sciences the following planning problem arises. A training set of
trajectories representing the system state, the actions and the
instantaneous rewards at successive discrete instants is known and a
decision policy leading to high cumulative rewards has to be found. In
the first part of this talk, I will present some batch mode
reinforcement learning algorithms for solving these types of problems.
Afterwards, I will report on some recent work to build lower bounds on
the performance of a policy from trajectories. Finally, I will show
that this work has opened a frontier in batch mode reinforcement
learning whose final implications we cannot yet guess.