Scalable Hardware Architecture for Real-Time Dynamic Programming Applications

This paper introduces a novel architecture for performing the core computations required by dynamic programming (DP) techniques. The latter pertain to a vast range of applications that necessitate an optimal sequence of decisions to be issued. An underlying assumption is that a complete model of the environment is provided, whereby the dynamics are governed by a Markov decision process (MDP). Existing DP implementations have traditionally been realized in software. Here, we present a method for exploiting the data parallelism associated with computing both the value function and optimal action set. An optimal policy is obtained four orders of magnitude faster than traditional software-based schemes, establishing the viability of the approach for real-time applications