Abstract
On 20 and 21 January 2025, the GameTable COST Action’s Working Group 1 convened in London for a meeting focused on game-playing artificial intelligence (AI), including search algorithms, knowledge representation, and reinforcement learning. The first day featured participant presentations, while the second day was dedicated to plenary and small group discussions. A significant outcome was the identified need for accessible resources to bridge AI and cultural heritage research. This report provides a summary of the discussions and talks that took place during the meeting.
Introduction
On 20 and 21 January 2025, we held a Working Group (WG) 1 meeting for the GameTable COST Action CA22145 (Piette et al., 2024), in London. GameTable is an interdisciplinary network of researchers, studying tabletop games from perspectives such as artificial intelligence (AI), cultural heritage, archaeology, mathematics, and education. Within this network, the focus of WG1 is on AI techniques for automated playing of games (Soemers et al., 2024), such as search algorithms (Russell & Norvig, 2020) and reinforcement learning (Sutton & Barto, 2018). A total of 26 different people participated in the meeting, of which 20 were in-person, with the remaining six joining remotely.
The first day of the meeting was primarily used for presentations contributed by some of the participants. These are summarized in Section 2. On the second day, our focus shifted to plenary discussions, as well as discussions in smaller groups. Some of the key topics of discussion are summarized in Section 3. Section 4 elaborates in further detail on one of the key conclusions from the meeting: the need for easily accessible information that can help AI and cultural heritage researchers better understand and communicate with each other. Finally, Section 5 concludes this paper.
Contributed Talks
Imperfect-Information Games
The first session of the meeting focused on the challenges and advancements in imperfect-information games, a critical area in game AI research. This session provided a platform to discuss the complexities associated with reasoning under uncertainty and the development of general strategies for such games.
The first talk, titled “Belief Stochastic Game: A Model for Imperfect-Information Games with Known Position,” was presented by Achille Morenville (2024a, 2024b). Imperfect-information games present significant challenges for general game playing (GGP; Genesereth et al., 2005) agents, as conventional models, such as Extensive Form Games (von Neumann & Morgenstern, 1944), and more recent ones such as Factored-Observation Stochastic Games (Kovarík et al., 2022), require agents to construct and maintain estimates of the game state. This often results in game-specific solutions and the unintended incorporation of domain-specific knowledge, limiting the generalizability of such approaches. To overcome these limitations, Achille introduced the Belief Stochastic Game model, a novel framework that externalizes state estimation by shifting it from the agent to the game model itself. This allows agents to concentrate exclusively on strategy development rather than on complex state inference. By exploiting common structural patterns in many imperfect-information games, this approach enhances the adaptability of AI agents, enabling them to generalize more effectively across diverse game environments.
The second talk, titled “Finding Portfolios of Opponent Strategies in Large Imperfect Information Games,” was presented by Drabent et al. (2024). She explored the use of strategy portfolios to improve decision-making in complex games where computing Nash equilibrium is infeasible. Instead of considering all possible strategies, agents can construct portfolios of opponent strategies to optimize their own play, either by minimizing worst-case losses (strategy optimization) or exploiting opponents (opponent exploitation). By restricting the game to a selected portfolio, computational efficiency is improved while maintaining strategic depth. She introduced methods such as Gradient Clustering Transformations and Regularized Nash Dynamics to refine portfolio selection, ensuring adaptability across large game spaces. Experimental results demonstrated the effectiveness of these approaches, but open questions remain on computing optimal mixed pessimistic portfolios. The findings highlight portfolio-based optimization as a promising alternative to exhaustive game-solving techniques.
The last talk of this session, titled “Adapting to opponents in large imperfect information games,” was presented by Milec et al. (2024, 2025). He explored methods for improving AI adaptability in complex game environments where computing the exact best response is infeasible. He discussed depth-limited approaches using value function approximation and heuristics such as DeepStack (Moravčík et al., 2017) to estimate game values beyond computational limits. To enhance robustness, he introduced Worst-Case ModelMix, which blends opponent modeling with worst-case planning to mitigate errors in strategy adaptation. The talk also examined agent evaluation techniques, comparing traditional head-to-head testing with exploitability metrics, which provide a more general measure of an AI’s adaptability. The talk concluded with open questions on adaptation in general-sum games and evaluating AI performance in Stratego and Dark Chess, highlighting ongoing challenges in imperfect-information game AI.
How Humans and AI Experience Game Playing
The second session was dedicated to how humans and AI experience game playing. GameTable members presented studies and projects on how AI models learn from experience, adapt to new challenges, and sometimes even exhibit behavior that mimics human intuition. Comparisons were drawn between human problem-solving approaches and the pattern recognition abilities of AI, highlighting both the strengths and limitations of each.
The first talk on this topic was titled “Quantifying tabletop games with AI—can we transfer anything to human experience?” and was given by James Goodman (2021a, 2021b, 2024, 2025). This presentation focused on quantifying tabletop games using AI-driven metrics to analyze game characteristics such as difficulty, randomness, and skill depth. His research aims to develop game fingerprints, which are distributions of optimized AI parameters that help categorize games based on computational play. Using Monte Carlo Tree Search (MCTS; Browne et al., 2012; Świechowski et al., 2022) optimization, he explored how games can be mapped into a structured landscape, allowing for comparative analysis across different titles. The talk covered several key topics, including game landscapes, measuring game difficulty using skill traces derived from AI performance, and assessing randomness by analyzing game outcomes under controlled conditions. While some findings aligned with human intuition—such as hidden information affecting strategy in Love Letter—others, such as the reported high skill depth of Sushi Go, raised questions about the transferability of AI-based quantifications to human gameplay. This talk concluded with reflections on agent limitations, highlighting the need for human validation to ensure AI-generated metrics truly reflect player experience.
The second and last talk on this topic was titled “Winning is not everything—Towards human-like agents for tabletop games” and was given by Aloïs Rautureau. He explored the development of human-like AI agents for tabletop games, emphasizing that winning is not the only goal in game-playing AI. While traditional GGP agents optimize for victory using techniques such as MCTS, this approach makes them unrealistic opponents for human players. The talk examined how human likeness can be defined and measured, incorporating insights from cognitive science, psychology, and AI research in other fields (e.g., chatbots and non-player character behavior in video games). A two-system thinking model was introduced, where System 1 represents intuitive pattern recognition and System 2 involves deeper analytical reasoning. A proposed framework for human-like agents in GGP integrates these systems by filtering out intuitively bad moves and using MCTS only when needed. Initial implementations for Renju demonstrated promising results, with ongoing work aiming to refine the model using inverse reinforcement learning (Ng & Russell, 2000; Russell, 1998) to infer human motivations. The talk concluded with future research directions, including integrating human-like AI into Ludii (Piette et al., 2020), exploring its role in ancient game reconstruction (e.g., Browne et al., 2019, 2022; Crist et al., 2024), and investigating AI models for cheating behavior in games.
Generalization and Explainability
The last talk session was dedicated to Generalization and Explainability. The session also discussed recent breakthroughs and ongoing debates on how best to balance model complexity with interpretability.
The first presentation, titled “Explainability of Board Game Agents,” was delivered by Manuel Eberhardinger and focused on the findings of a Short-Term Scientific Mission conducted during the first grant period of the GameTable COST Action. The talk focused on improving the explainability of AI agents in board games, particularly by using decision trees and genetic programming to generate human-interpretable explanations of AI decisions. The motivation behind this work is to address the black-box nature of game-playing AI, such as MCTS and reinforcement learning agents, which makes understanding their strategic choices difficult. The research aims to extract state-action features (Soemers et al., 2023a, 2023b) without relying on expert policies or neural network logits, allowing for a more transparent decision-making process. The proposed method involves genetic programming to discover board game features and training decision trees that predict AI actions based on these features. Initial evaluations, using AlphaZero-trained agents, demonstrated that while decision trees can approximate AI strategies, they often fail to generalize correctly, leading to brittle decision-making. Future directions include testing the framework on simpler games such as Tic-Tac-Toe, refining the feature selection process, and developing a learnable domain-specific language to improve the explainability and robustness of AI-driven board game strategies.
The final talk, titled “Games and out-of-distribution generalisation” was given by Samothrakis et al. (2024) and Soemers et al. (2025). It explored out-of-distribution (OOD) generalization in AI and game-playing agents, questioning whether current AI approaches can truly generalize beyond their training environments. The talk reviewed the evolution of game AI from rule-based systems to deep reinforcement learning, highlighting the “Bitter Lesson”—that AI progress tends to come from scaling computation rather than hand-crafted knowledge (Sutton, 2019). However, despite advances such as AlphaZero, current AI models still struggle with sample efficiency and OOD generalization, requiring vast amounts of training data to adapt to unseen scenarios. The presentation introduced key OOD challenges, including systematicity, productivity, and substitutivity, which relate to how AI recognizes and applies learned patterns in novel situations. Samothrakis argued that procedural content generation alone is insufficient for true generalization, and instead, new approaches combining neural networks with symbolic reasoning might be needed. He concluded with open research questions on integrating deep learning with first-order logic and developing more efficient algorithms that can generalize across a wide range of games and real-world tasks.
Discussions
In addition to contributed talks, we reserved a substantial amount of time for more open-ended discussions. One of the main topics of discussion was how best to facilitate further communication and collaboration between researchers studying games from the AI perspective on the one hand, and the cultural heritage perspective on the other hand. We dedicate an entirely separate section—Section 4—to this topic. The following other topics emerged as key topics for further consideration in the research community:
Human-like AI: how can we implement AI algorithms that play tabletop games such as humans do, such that any measures we collect from simulations accurately estimate the experience that humans would have had to play that ruleset? How can we make AI follow not only the explicit rules defined in games’ rulesets—and maybe deliberately not follow them in plausible ways—but also follow social etiquette rules (e.g., avoid moving back and forth indefinitely)? Another interesting factor is the thinking time used when playing: depending on culture and social context, certain amounts of thinking time may or may not be considered socially acceptable (or even allowed by tournament rules), but humans and AI players tend to be affected by time in different ways. Explainability in search and reinforcement learning for game playing, and AI systems that can give advice or recommendations as to how to play to humans. How can we implement effective frameworks and benchmarks to facilitate research in the combination of GGP with imperfect-information games? Which tabletop games, if any, remain major challenges where AI cannot yet reach superhuman levels of playing strength? How can we improve methodologies used for benchmarking and evaluating different AI algorithms for game playing? Development of AI that can effectively collaborate with humans or other AI agents in collaborative games. What role can games play in benchmarking for Artificial General Intelligence? How can we use AI and games in education? How can we effectively share teaching materials, and in general collaborate in educational activities across universities, within the game AI research community?
Combining AI and Cultural Heritage for Tabletop Games Research
One of GameTable’s overarching aims is to bring together experts in AI and the cultural heritage of games to identify and test new methodologies for approaching past ludic activity (Piette et al., 2024). To further this goal, four members (Walter Crist, Summer Courts, Tim Penn, and Ilaria Truzzi) of WG2—“Cultural Heritage of Games”—attended the meeting to identify and discuss viable avenues for future research in this area. A key theme to emerge from conversations between the WG1 and WG2 members at this meeting is that experts in AI and experts in the cultural heritage of games work within highly divergent research traditions and frameworks. This divergence underscores the need for open dialogue to foster meaningful, collaborative research. Given that the application of AI to historical games remains a nascent field, participants agreed that developing well-defined case studies would be an effective capacity-building strategy to bridge these disciplinary gaps.
Very few concrete case studies that apply AI to answer questions about historical games have been published so far (Browne, 2023; Crist et al., 2024; Donkers et al., 2000). For the field to advance, continued collaboration between WG1 and WG2 members is essential in developing viable AI-based approaches to studying past games. This requires formulating specific research questions grounded in the distinct characteristics of specific traditional games. A key challenge is determining which metrics of traditional games can be reasonably calculated using AI and what types of research questions these methods can address to create useful new insights for scholars working on historical games. One promising area identified during the meeting is games that rely on chance—particularly those involving randomization devices such as dice or knucklebones (astragals). During discussions in the meeting, participants identified several avenues for future research, perhaps to be explored as part of a journal special issue on AI and historical games.
Given a hypothesized ruleset for a game, AI-driven players can be used to simulate play at a large scale (e.g., running hundreds or thousands or more of simulated plays) and perform quantitative analysis at a level that would be infeasible to do with human playtesting. Essentially any quantity of interest that can be given a clear mathematical definition can be measured and analyzed from such simulations. Examples include:
Duration (in number of moves or turns) per game. If games consistently and easily end in an extremely short time (e.g., the first player can win immediately), the evaluated ruleset is not plausible (Browne, 2023). Various estimators of the “quality” of a game can be measured, following the intuition that rulesets that people enjoy playing are more likely to have been played than low-quality rulesets (Browne, 2018; Browne et al., 2019; Crist et al., 2024). Duration could again be one factor in this, if we assume that games are considered better or more fun if they take neither too long nor too short to complete. However, care should be taken to account for differences between cultures and social contexts within games played, as these can affect how much time is considered too short or long. Other factors could include balance (does each player have a fair chance at winning), skill depth (is there room for different levels of skill expression; Browne, 2022; Goodman et al., 2024), and more (Browne, 2009; Kowalski & Szykuła, 2016). Usage of game equipment. If certain parts of a game’s equipment (e.g., certain pieces or certain parts of the board) see substantially more use than others in simulated play, this could be correlated to signs of usage visible in the archaeological material. The impact of using different sources of randomness (e.g., different types of dice, as mentioned previously as a possible case study from cultural heritage research) on game outcomes and aspects such as the room for skill expression may be estimated from AI-driven playtesting (Goodman et al., 2025). Measures as described in the previous two points may be used to generate plausible explanations as to why certain changes in rules between closely related games may have been introduced. Differences in rules between games can be correlated to measures of quality, measures of balance between randomness and skill, and so on. When our knowledge of rulesets of an ancient game is incomplete (i.e., we know some parts of the rules, but not the complete rules), we may attempt to automatically fill in the missing parts. This may, for instance, be done by copying relevant rules from games that are closely related in terms of, for example, cultures or social contexts in which the games were played. Such a process could procedurally generate a wide variety of hypothesized rulesets, each of which in turn could be evaluated for plausibility as described previously. If using solely AI-driven evaluations by themselves is not sufficiently reliable, an alternative approach can be to use a combination of AI-driven and human playtesting. AI-driven playtesting can filter a wide variety of hypothesized rulesets down to a smaller set, which may then be further tested by human players. This may require the development of extensive tools that enable convenient, extensive playtesting of arbitrary (potentially procedurally generated) tabletop games, ideally in an online interface. Given an exhaustive database containing detailed information about games played throughout history, including extensive data on what is known about their rules, the social contexts in which they were placed, and any information concerning the geographical locations and periods of time in which they were played that may be derived from archaeological evidence (e.g., Crist et al., 2022), data science techniques may be used to generate plausible ways of filling in gaps. However, the sparsity of existing data remains a concern for this idea.
This paper summarized the key topics of discussion and outcomes from the January 2025 meeting of WG1 of the GameTable COST Action (Piette et al., 2024). There were numerous discussions and talks surrounding various aspects of AI research for tabletop games, the use of tabletop games for AI research, and how to drive the field forward. In this report, we placed particular emphasis on the matter of how best to facilitate interdisciplinary research between AI researchers and cultural heritage researchers. We identified a need to provide examples of (1) research questions that are of interest to researchers studying games from a cultural heritage perspective, and (2) techniques and methods that AI researchers could contribute to help answer such research questions. Such lists of examples should help researchers across the different disciplines to more easily and effectively communicate with each other. A first attempt at fulfilling this need is included in this report.
Footnotes
Acknowledgments
This article is based upon work from COST Action CA22145—GameTable, supported by COST (European Cooperation in Science and Technology). We thank all the participants who attended and contributed to our meeting.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.