Regresar

Shoot-Out: Exploring HyperNEAT for an optimal Final-Third Approach in Robocup-2D Soccer

Abstract:

Past research have propositioned HyperNEAT to be suitable for domains that express spatial regularities. The regularities facilitate an indirect encoding scheme that expresses the topology of the decision-making Artificial Neural Network (ANN). The Robocup-2D soccer sub-task of Keepaway is one such domain, where HyperNEAT has shown not only to be effective at solving Keepaway, but also task transferable by scale using back the same CPPN that encodes the ANN topology. In this paper, we explore whether HyperNEAT could not only train a team of keepers to perform Keepaway at the opponents' half, but also for each keeper to consider shooting towards a well-guarded goal with an opposing taker scripted as the goalkeeper. In short, the expanded task is termed 'Shoot-Out'. The paper presents the extensions to Verbancsis's HyperNEAT-BEV framework for Shoot-Out, and a discussion of the results obtained against runs from a policy where keepers take random actions (Random), and also against a policy of keepers with scripted behaviors (Hand-Coded).