Workshop on computational intelligence and games 2024

The Workshop on Computational Intelligence and Games (CIG2024) was held on April 13 2024 at the Southern University of Science and Technology (SUSTech) in Shenzhen, China and it was chaired and organized by Dr Jialin Liu. The workshop was supported by the SINO-MALTA project “Online Procedural Content Generation via Multi-objective Optimisation and Learning” (OPtiMaL).1

The aim of the workshop was to demonstrate the various roles of artificial intelligence (AI) across different fields, particularly the ways AI approaches and techniques that are widely studied in the computer games research community are being leveraged to advance and help other research fields and applications (such as creative design tasks, science and discovery). For that purpose we invited eight world-class keynote speakers representing both academic research and industrial innovation in AI and games. A panel session followed the eight keynote talks. The keynotes were grouped into four sessions based on the talk topic. Those sessions were named as follows: Human-centered AI, AI for Science & Industry, Search Methods and Their Applications, and AI for Design. Moreover, two student poster sessions were held throughout the day. Over 50 participants attended this event, including not only researchers, students and engineers from academia and industry, but also high-school teachers and teenagers. We briefly report on the activities of the workshop below.

Session I: Human-centered AI

In the first talk Human Feedback via Games, Georgios N. Yannakakis (University of Malta, Malta) presented the importance of understanding human feedback for enhancing player experience, and incorporating human feedback into the design of games and creative AI algorithms through a series of milestone research studies using car racing games, first-person shooter games and platformer games.

In the second talk What Comes After Human-Centered AI? Design Insights from Games, Well-being Apps, and Art, Jichen Zhu (IT University of Copenhagen, Denmark), highlighted the position of player-AI interaction as a multidisciplinary field of study focusing on the design and interaction between humans and the AI in the context of play, presented their recent work on designing AI-based experiences in various contexts, particularly Player-AI Interaction: What Neural Network Games Reveal About AI as Play, and proposed research directions toward designing next-generation human-AI interaction.

Session II: AI for Science & Industry

In the talk The Promise of Multi-objective Sequence Learning, Mike Preuss (Universiteit Leiden, Netherlands) presented how Monte Carlo tree search can be used in a multi-objective context to solve different problems in two distinct fields, computer games and chemistry syntheses. The presented studies not only serve as a remarkable example of successful interdisciplinary research, but also show the different roles of AI approaches in decision-making, science and discovery.

Our industry speaker Alexis Rolland (Ubisoft La Forge, China) presented Pioneering 2D Image Generation for AAA Games. He first presented the role and vision of La Forge at Ubisoft, and then introduced the evolution of image generation models, examined the impact of image generation in game development workflows and in the game industry with examples, and finally highlighted the Ubisoft’s approach and strategy to leverage image generation in a responsible fashion.

Session III: Search Methods and Their Applications

The first presentation after lunch was AI for Tabletop Games given by Diego Perez-Liebana (Queen Mary University of London, UK). Different from traditional board games, tabletop games provide more challenges (such as hidden information, random events, long term vs. short term planning, and gambling). Diego Perez-Liebana presented their latest work on playing and designing tabletop games, variants of Monte Carlo tree search, heuristics, visualisation of game spaces and AI-assisted game design. Future research directions using their Tabletop Games frameworks (TAG and PyTAG) were also discussed.

Then, in the talk What’s Next for Adversarial Search Techniques?, Mark Winands (Maastricht University, NL) reviewed some inspiring success cases of adversarial search techniques such as alpha-beta search and Monte-Carlo tree search in board games and general game playing, and then discussed how those techniques can be applied to real-world applications. Their recent attempts such as explainable search, self-adaptive (or generative) search engines, and integrating search with quantum computing were also presented.

Session IV: AI for Design

In the presentation Empowering Game Designers with Automatic Playtesting, Raluca D. Gaina (Tabletop R&D & Queen Mary University of London, UK), being a faculty and a company co-founder at the same time, bridged academic research and industrial innovation. The complexity of modern tabletop games lead to an increase in time and effort spent by designers developing and playtesting new games, posing a significant challenge to independent designers and small companies. This talk presented how Tabletop R&D addressed these issues using the latest AI technologies and digital twins of tabletop games.

In the last talk of the workshop Text2Game: Text-conditioned Generation of Game Levels and Other Game Things, Julian Togelius (New York University, US) presented their latest progresses in generating game content through Large Language Models (LLMs) and discussed a number of strategies for training text-conditioned game content generators (such as fine-tuning, bootstrapping, and generator distillation) using examples, ranging from tile-based 2D game levels to 3D Minecraft objects and game rules.

Panel: LLMs and Games

The workshop ended with the panel session LLMs and Games, co-chaired by Julian Togelius and Jialin Liu. Speakers and audience were invited to answer five questions raised by students and conduct open discussions. The questions are listed below for further discussion.

What are potential directions for LLMs and games?

How can junior researchers survive without enough GPUs, meeting the impact of LLMs?

Are LLMs trustworthy?

How do variations in the amount and diversity of training data influence LLM performance between tasks or games?

How can we ensure that LLMs sustainably learn and evolve between games without experiencing catastrophic forgetting or performance decay over time?

Student Poster Sessions

Two poster sessions were organised during coffee breaks between the two keynote sessions in the morning and afternoon, in which volunteer students from the Learning and Optimisation in Games (LOG) research group of SUSTech presented their recent research publications. Below we list the presented posters and corresponding presenters.

Negatively Correlated Ensemble Reinforcement Learning for Online Diverse Game Level Generation by Ziqi Wang.

Reinforcement Learning with Dual-observation for General Video Game Playing by Chengpeng Hu.

Exploring a Computer Embroidery Swatchbook in a University Classroom by Yuchen Li.

Measuring Diversity of Game Scenarios by Yuchen Li.

Fun As Moderate Divergence: Evaluating Experience-Driven PCG via RL by Ziqi Wang and Yuchen Li.

Mitigating Unfairness via Evolutionary Multi-objective Ensemble Learning by Qingquan Zhang.

fSDE: Efficient Evolutionary Optimisation for Many-objective Aero-engine Calibration by Qingquan Zhang.

All participants enjoyed the workshop and interacted actively during the keynotes, panel, poster sessions, coffee breaks and lunch. The workshop received positive feedback from all speakers and its diverse audience.