Poster ID T4 Incremental Natural Actor-Critic Algorithms

'

$

S. Bhatnagar , R. Sutton , M. Ghavamzadeh , & M. Lee · We developed four new actor-critic algorithms and provided their convergence analysis. · Our algorithms are based on conventional & natural gradients. · Our results extend prior convergence results for actor-critic methods by Konda and Tsitsiklis by ­ using temporal difference learning in both actor & critic, ­ incorporating natural gradients. · Our results extend prior empirical studies of natural actor-critic methods by Peters et al. by ­ providing the first convergence proofs,

­ presenting the first fully incremental algorithms. & %   Indian Institute of Science, Bangalore University of Alberta, Canada