How to solve large scale deterministic games with mean payoff by policy iteration

Min-max functions are dynamic programming operators of zero-sum deterministic games with finite state and action spaces. The problem of computing the linear growth rate of the orbits (cycle-time) of a min-max function, which is equivalent to computing the value of a deterministic game with mean payoff, arises in the performance analysis of discrete event systems. We present here an improved version of the policy iteration algorithm given by Gaubert and Gunawardena in 1998 to compute the cycle-time of a min-max functions. The improvement consists of a fast evaluation of the spectral projector which is adapted to the case of large sparse graphs. We present detailed numerical experiments, both on randomly generated instances, and on concrete examples, indicating that the algorithm is experimentally fast. Categories and Subject Descriptors G.2.2 [Discrete Mathematics]: Graph Theory; G.4 [Mathematical software]: Algorithm design and analysis General Terms Algorithms, Performance Keywords...

• Type: Conference
• Year: 2006
• Where: VALUETOOLS
• Authors: Vishesh Dhingra, Stephane Gaubert
• Keywords: Deterministic Game, Hardware, Min-max Functions, Policy Iteration, VALUETOOLS 2006