Seth Cooper’s research while affiliated with Northeastern University and other places

WaveFunctionCollapse (WFC) is a constraint-satisfaction-based approach to procedural content generation via machine learning (PCGML). It is relatively easy to implement and requires very little training data, making it a popular approach. Generated game levels are guaranteed to look locally similar totheir example tilemaps; however, local adjacency rules often fail to capture global solvability rules, potentially making many such levels unplayable. Existing approaches to improving the solvability of WFC-generated levels typically require adding additional game-specific information in the form of global constraints, substantially increasing the complexity and time required for setup. The purpose of this work is to explore whether using level solutions as training data can allow WFC to learn solvability constraints and game mechanics. We have implemented a novel space-time approach that uses three-dimensional space-time blocks representing solutions to 2D levels as both input and output. Experiments using this method show that space-time WFC is capable of demonstrating localized game mechanics and creating small playable levels with given solutions. However, levels are slow to generate, and some high-level constraints are still not captured.

Creating well-balanced combat encounters can be a difficult task for Game Managers (GMs) in tabletop games such as Dungeons and Dragons (DnD). This work uses a simulation environment to generate new sets of DnD encounters that can be optimized for both difficulty and balance among player contributions. Encounters are evaluated using simulated games that can either be run probabilistically (using dice rolls) or with deterministic expected outcomes. While the expected approach allows game outcomes to be simulated substantially faster and is a good estimate of difficulty, it is a less reliable measure of balance. A genetic algorithm was used to generate encounters that meet the desired difficulty and where all players are needed for success.

Procedurally generated game levels should be completable. The representation used for levels and game mechanics impacts the types of games for which different techniques can be applied. Previous work used a constraint solving approach to simultaneously generate levels with example playthroughs, showing they can be completed using the game's mechanics. However, that work used 2D grid-based rewrite rules. In this work, we extend previous approaches by representing levels as more general graphs, and game mechanics as rewrites on node and edge labels of subgraphs. Using this approach, graph-based levels with playthroughs are generated. We describe the approach and demonstrate its application in some games with graph-based levels.

2D tile-based levels are a common format for video game research, particularly in procedural content generation. Often, tiles are represented as text characters. Here, we describe level2image, a utility that provides a flexible means for converting such levels into a variety of image formats. It can use text tiles or substitute image tiles or backgrounds. There is support for various geometry overlays, such as paths through the level, areas of interest, or boundaries. The utility is a Python script, intended to reduce duplicated work creating such converters within the game research community.

Procedurally generated levels created by machine learning models can be unsolvable without further editing. Various methods have been developed to automatically repair these levels by enforcing hard constraints during the post-processing step. However, as levels increase in size, these constraint-based repairs become increasingly slow. This paper proposes using explainability methods to identify specific regions of a level that contribute to its unsolvability. By assigning higher weights to these regions, constraint-based solvers can prioritize these problematic areas, enabling more efficient repairs. Our results, tested across three games, demonstrate that this approach can help to repair procedurally generated levels faster.

Generative Adversarial Networks (GANs) are unsupervised models designed to learn and replicate a target distribution. The vanilla versions of these models can be extended to more controllable models. Conditional Generative Adversarial Networks (CGANs) extend vanilla GANs by conditioning both the generator and discriminator on some additional information (labels). Controllable models based on complementary learning, such as Rumi-GAN, have been introduced. Rumi-GANs leverage negative examples to enhance the generator's ability to learn positive examples. We evaluate the performance of two controllable GAN variants, CGAN and Rumi-GAN, in generating game levels targeting specific constraints of interest: playability and controllability. This evaluation is conducted under two scenarios: with and without the inclusion of negative examples. The goal is to determine whether incorporating negative examples helps the GAN models avoid generating undesirable outputs. Our findings highlight the strengths and weaknesses of each method in enforcing the generation of specific conditions when generating outputs based on given positive and negative examples.

Determining if and how the difficulty of algorithmic puzzle solvers is related to the difficulty and enjoyment for human players is a challenging task. In this work, we explored this relationship using logic grid puzzles. We used an algorithmic solver to estimate the difficulty of the puzzles by capturing the number of ``solver loops'' through the algorithm. This characteristic was used to generate and evaluate a set of puzzles of varying algorithmic difficulty using constrained MAP-Elites. Then, we ran a user study to gather information on the player experience of these puzzles. We tested the relationship between solver loops and player experience on generated puzzles and found that the number of solver loops is statistically significantly correlated with subjective perception of difficulty and borderline statistically significantly correlated with puzzle correctness.

Generative Adversarial Networks (GANs) are unsupervised models designed to learn and replicate a target distribution. The vanilla versions of these models can be extended to more controllable models. Conditional Generative Adversarial Networks (CGANs) extend vanilla GANs by conditioning both the generator and discriminator on some additional information (labels). Controllable models based on complementary learning, such as Rumi-GAN, have been introduced. Rumi-GANs leverage negative examples to enhance the generator's ability to learn positive examples. We evaluate the performance of two controllable GAN variants, CGAN and Rumi-GAN, in generating game levels targeting specific constraints of interest: playability and controllability. This evaluation is conducted under two scenarios: with and without the inclusion of negative examples. The goal is to determine whether incorporating negative examples helps the GAN models avoid generating undesirable outputs. Our findings highlight the strengths and weaknesses of each method in enforcing the generation of specific conditions when generating outputs based on given positive and negative examples.

Procedurally generated levels created by machine learning models can be unsolvable without further editing. Various methods have been developed to automatically repair these levels by enforcing hard constraints during the post-processing step. However, as levels increase in size, these constraint-based repairs become increasingly slow. This paper proposes using explainability methods to identify specific regions of a level that contribute to its unsolvability. By assigning higher weights to these regions, constraint-based solvers can prioritize these problematic areas, enabling more efficient repairs. Our results, tested across three games, demonstrate that this approach can help to repair procedurally generated levels faster.