GECCO '19- Proceedings of the Genetic and Evolutionary Computation Conference

Quality diversity (QD) algorithms such as MAP-Elites have emerged as a powerful alternative to traditional single-objective optimization methods. They were initially applied to evolutionary robotics problems such as locomotion and maze navigation, but have yet to see widespread application. We argue that these algorithms are perfectly suited to the rich domain of video games, which contains many relevant problems with a multitude of successful strategies and often also multiple dimensions along which solutions can vary.

This paper introduces a novel modification of the MAP-Elites algorithm called MAP-Elites with Sliding Boundaries (MESB) and applies it to the design and rebalancing of Hearthstone, a popular collectible card game chosen for its number of multidimensional behavior features relevant to particular styles of play. To avoid overpopulating cells with conflated behaviors, MESB slides the boundaries of cells based on the distribution of evolved individuals. Experiments in this paper demonstrate the performance of MESB in Hearthstone. Results suggest MESB finds diverse ways of playing the game well along the selected behavioral dimensions. Further analysis of the evolved strategies reveals common patterns that recur across behavioral dimensions and explores how MESB can help rebalance the game.

This paper provides a detailed investigation of using the Kullback-Leibler (KL) Divergence as a way to compare and analyse game-levels, and hence to use the measure as the objective function of an evolutionary algorithm to evolve new levels. We describe the benefits of its asymmetry for level analysis and demonstrate how (not surprisingly) the quality of the results depends on the features used. Here we use tile-patterns of various sizes as features.

When using the measure for evolution-based level generation, we demonstrate that the choice of variation operator is critical in order to provide an efficient search process, and introduce a novel convolutional mutation operator to facilitate this. We compare the results with alternative generators, including evolving in the latent space of generative adversarial networks, and Wave Function Collapse. The results clearly show the proposed method to provide competitive performance, providing reasonable quality results with very fast training and reasonably fast generation.

The capacity of genetic programming (GP) to evolve a 'hero' character in the Dota 2 video game is investigated. A reinforcement learning context is assumed in which the only input is a 320-dimensional state vector and performance is expressed in terms of kills and net worth. Minimal assumptions are made to initialize the GP game playing agents - evolution from a tabula rasa starting point - implying that: 1) the instruction set is not task specific; 2) end of game performance feedback reflects quantitive properties a player experiences; 3) no attempt is made to impart game specific knowledge into GP, such as heuristics for improving navigation, minimizing partial observability, improving team work or prioritizing the protection of specific strategically important structures. In short, GP has to actively develop its own strategies for all aspects of the game. We are able to demonstrate competitive play with the built in game opponents assuming 1-on-1 competitions using the 'Shadow Fiend' hero. The single most important contributing factor to this result is the provision of external memory to GP. Without this, the resulting Dota 2 bots are not able to identify strategies that match those of the built-in game bot.

In this paper we investigate the effect of embedding different levels of musical knowledge in the virtual machine (VM) architectures and phenotype representations of an algorithmic music composition system. We examine two separate instruction sets for a linear genetic programming framework that differ in their knowledge of musical structure: one a Turing-complete register machine, unaware of the nature of its output; the other a domain-specific language tailored to operations typically employed in the composition process. Our phenotype, the output of the VM, is rendered as a musical model comprising a sequence of notes represented by duration and pitch. We compare three different pitch schemes with differing embedded knowledge of tonal concepts, such as key and mode.

To derive a fitness metric, we extract musical features from a corpus of Hungarian folk songs in the form of n-grams and entropy values. Fitness is assessed by extracting those same attributes from the phenotype and finding the maximal similarity with representative corpus features.

With two different VM architectures and three pitch schemes, we present and compare results from a total of six configurations, and analyze whether the domain-specific knowledge impacts the results and the rate of convergence in a beneficial manner.