Multi-objective model to address planning in a RTS game

Abstract

Real-time strategy (RTS) games hold many challenges in the creation of a game AI. One of those challenges is creating an effective plan for a given context. Competitive game AIs have struggled to adapt and create good plans to counter the opponent strategy. In this paper, a new scheduling model is proposed regarding planning problems on RTS games. This problem consists of solving a multi-objective model that satisfies constraints based on a given system from an initial state to achieve an adequate strategy. The system contains resources, tasks and cyclic events that translate the game into an instance of the problem. The initial state contains information about the state of the resources, uncompleted tasks and on-going events. The strategy defines which resources to maximize or minimize (if any), and which constraints are applied to the resources, as well as to the project horizon. Two algorithms for creating solutions are introduced, one based upon weights for different needs and another based on the ant-colony optimizer. Two optimizers — NSGA-II, and multi-objective ACO — are compared on instances based on real RTS problems.