Portal de Periódicos da CAPES

Guest Editorial: Special Issue on Evolutionary Computation for Games

2023; Institute of Electrical and Electronics Engineers; Volume: 15; Issue: 1 Linguagem: Inglês

10.1109/tg.2022.3225730

2475-1510

Jacob Schrum, Jialin Liu, Cameron Browne, Anikó Ekárt, Marcus Gallagher,

Sports Analytics and Performance

T HE application of evolutionary computation (EC) [1] is widespread because the core method is very general.If you have a way to represent candidate solutions, modify candidate solutions and evaluate candidate solutions, then you can apply EC.No complex formalism or theoretical framework is needed, which makes EC easy to apply, and lowers the barrier for entry by new researchers.This generality is particularly useful for games, because there are many different aspects of games that can benefit from searching for better solutions.What is more, EC is also a powerful method with many benefits over traditional optimization methods.Because evolution is population based, multiple solutions to a problem can be produced rather than just one.Having a population of solutions is extremely useful in multimodal problems, multiobjective problems, problems with nonstationary fitness landscapes, and situations where a diversity of different solutions is desired.All of these situations are common in games: good games typically allow agents to solve problems in a variety of different ways (multimodal), different players/agents may have preferences for different objectives (multiobjective), the value of a given evolved artifact or agent strategy may change in response to player actions in an online game (non-stationary), and many games focus on collecting a wide range of different novel artifacts (diversity).Studies exploring various aspects of games with EC are abundant.There have been numerous examples of evolution applied to the discovery of agent behavior in games, from board games, such as Checkers [2] and Chess [3], to classic arcade games, such as Ms. , Super Mario Bros. [5], and various Atari games [6], all the way up to modern games including first-person shooters [7] and car racing [8].Evolution has also proven effective as a form of procedural content generation (PCG) [9], having been used to generate maps and levels [10], artistic content [11], and weapons [12].Evolution also blends well with many other techniques that are relevant in the playing and designing of games [13], [14].Machine learning, particularly with deep neural networks, has dominated many areas of AI research in recent years.Generative adversarial networks and variational autoencoders provide new ways to induce a latent space of possibilities that give rise to novel artistic designs, and these vector spaces can be searched by EC.In the realm of agent control, many successful evolutionary approaches are actually a combination of evolution with some Digital Object