Creating AI Characters for Fighting Games Using Genetic Programming

Abstract

Abstract-This paper proposes a character generation approach for the M.U.G.E.N. fighting game that can create engaging AI characters using a computationally cheap process without the intervention of the expert developer. The approach uses a Genetic Programming algorithm that refines randomly generated character strategies into better ones using tournament selection.The generated AI characters were tested by twenty-seven human players and were rated according to results, perceived difficulty and how engaging the gameplay was. The main advantages of this procedure are that no prior knowledge of how to code the strategies of the AI character is needed and there is no need to interact with the internal code of the game. In addition, the procedure is capable of creating a wide diversity of players with different strategic skills, which could be potentially used as a starting point to a further adaptive process.

Interpretable Utility-based Models Applied to the FightingICE Platform

“…This would certainly need a lot of engineering, requiring us to modify the platform. Finally, we can cite the work of Martínez-Arellano et al [10], using genetic algorithm to generate enjoyable strategies to play against in the MUGEN engine. Like this work, we also use a genetic algorithm to learn models of our ICNs (although this is not the main point of our paper).…”

Coping with opponents: multi-objective evolutionary neural networks for fighting games

“…On the other hand, approaches exist that do not require any domain-specific knowledge. Among these are genetic programming [35], neuroevolution [36], and other evolutionary approaches, and Monte Carlo tree search [3]. The first two are based on the biological process of evolution and aim at iteratively creating solutions of higher performance by mutating and combining existing ones.…”

“…The first two are based on the biological process of evolution and aim at iteratively creating solutions of higher performance by mutating and combining existing ones. They mainly differ in the representation of their solutions: While genetic programming typically encodes candidate solutions as binary trees, where nodes are functions or logical operators and leaves represent the data being processed [35], neuroevolution evolves artificial neural networks as described in this paper. Note that Martínez-Arellano et al [35] replaced the binary tree by a sequential encoding of possible actions.…”

“…They mainly differ in the representation of their solutions: While genetic programming typically encodes candidate solutions as binary trees, where nodes are functions or logical operators and leaves represent the data being processed [35], neuroevolution evolves artificial neural networks as described in this paper. Note that Martínez-Arellano et al [35] replaced the binary tree by a sequential encoding of possible actions. On the other hand, Monte Carlo tree search iteratively builds a search tree, where every path from the root to a leave is a possible sequence of actions.…”

Application of Retrograde Analysis on Fighting Games

“…These types of algorithms rely on dynamic scripting [23,24] to learn the optimal state with which to execute each individual policy. More complicated algorithms such as genetic algorithms [19], neural networks [20,21,22], and hierarchical reward architecture [26] have all been implemented on the fighting ICE framework and have learned strategies that are stronger than the "default" rule based AI.…”