Large Scale Feature Construction for Atari Games
Mayank Daswani (ANU)
ARTIFICIAL INTELLIGENCE SEMINARDATE: 2013-10-30
TIME: 11:30:00 - 12:00:00
LOCATION: RSISE Seminar Room, ground floor, building 115, cnr. North and Daley Roads, ANU
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
Traditional function approximation techniques have large computation and memory costs for very large feature sets. A particularly interesting class of algorithms impose a sparsity constraint on the regulariser for the weight vector but starting with a very large feature vector still has computational and memory disadvantages. In the general reinforcement learning setting where features are also over the history space, this can be impractical. The recently introduced Arcade Learning Environment (ALE) is a new testbed for reinforcement learning algorithms. The environments in this setting are games made for humans which can be relatively complex, but due to the space/processing limits of the ATARI console, still feasible for current RL techniques. When dealing with ATARI games the number of features used over the visual field are themselves very large. This talk will discuss our recently introduced model-free feature reinforcement learning algorithm which incrementally constructs features, along with improvements and ongoing work necessary for scaling to the class of ATARI games.
BIO:
