An Improved Rolling Horizon Evolution Algorithm with Shift Buffer for General Game Playing

Abstract

General Game Playing (GGP) is the design of artificial intelligence programs to play more than one game. Here, one of the most famous GGP frameworks, The General Video Game AI Competition (GVGAI) Framework, is used in order to design controllers for Atari 2600 inspired games. Recent advancements in the literature of GVGAI showed that the Rolling Horizon Evolution Algorithm (RHEA) is competitive when compared to other methods, encouraging the use and the research by improvements for this method. The use of a 1-Step-Look-ahead approach and a Redundant Action Avoidance policy during the creation of new individuals are proposed in this paper. The 1-step-look-ahead technique improves the action selection after the shift of the individual in RHEA with the shift buffer enhancement (RHEA-SB), and the redundant action avoidance policy decreases the chance of spatial redundant actions within the individual. Also, a parameter analysis of RHEA-SB is performed here, where different values of population size, depth of simulations, and number of individuals that remains in the population are evaluated. Results show that using 1-Step-Look-ahead and a redundant action avoidance policy improves the quality of the solutions found when compared to the original algorithm.