General Video Game AI: a Multi-Track Framework for Evaluating Agents, Games and Content Generation Algorithms (1802.10363v4)

Abstract: General Video Game Playing (GVGP) aims at designing an agent that is capable of playing multiple video games with no human intervention. In 2014, The General Video Game AI (GVGAI) competition framework was created and released with the purpose of providing researchers a common open-source and easy to use platform for testing their AI methods with potentially infinity of games created using Video Game Description Language (VGDL). The framework has been expanded into several tracks during the last few years to meet the demand of different research directions. The agents are required either to play multiple unknown games with or without access to game simulations, or to design new game levels or rules. This survey paper presents the VGDL, the GVGAI framework, existing tracks, and reviews the wide use of GVGAI framework in research, education and competitions five years after its birth. A future plan of framework improvements is also described.

• The paper introduces a multi-track framework leveraging VGDL to assess AI agents in planning, learning, and content generation.
• It details diverse competition tracks, including forward model planning and episodic learning, showcasing varied algorithmic performances.
• The study outlines future directions for robust generalization and dynamic adaptation in video game AI research.

Overview of "General Video Game AI: a Multi-Track Framework for Evaluating Agents, Games, and Content Generation Algorithms"

The paper "General Video Game AI: a Multi-Track Framework for Evaluating Agents, Games and Content Generation Algorithms" presents a comprehensive framework for the design and assessment of AI agents capable of general video game playing (GVGP). It discusses the General Video Game AI (GVGAI) framework, rooted in the objectives of fostering broader AI capabilities beyond specialization in single games. The necessity for such a framework is underscored by the historical tendency for agents to overfit to specific games, thereby undermining the advancement of general AI.

Core Components

The GVGAI framework utilizes the Video Game Description Language (VGDL), which facilitates the creation of a vast array of two-dimensional arcade-style games. The framework is unique in offering multiple competition tracks requiring agents to either play unknown games or generate content such as levels or rules. Each track poses distinct challenges—such as using a forward model for game simulation in the planning tracks or relying purely on episodic learning in the absence of such a model.

Competition Structure

The paper elaborates on various competition tracks:

1. Single-Player and Two-Player Planning Tracks: These tracks challenge agents to plan and predict outcomes using a forward model. Agents like OLETS and MCTS variants dominate these tracks with their ability to perform open loop and tree search optimizations.
2. Single-Player Learning Track: Here, agents lack access to a forward model and must learn through episodic interactions. Algorithms such as Q-Learning and SARSA have been employed, though with varying levels of success.
3. Procedural Content Generation (PCG) Tracks: These tracks involve generating game levels and rule sets, with approaches ranging from cellular automata to genetic algorithms being explored.

Analytical Discussion

The authors provide an in-depth analysis of different AI methodologies utilized in these competitions. They note that while tree search methods, particularly those enhancing MCTS with various strategies (e.g., biasing rollouts with domain knowledge or employing multi-objective optimization), have shown promise, results are often game-dependent. Less conventional approaches, such as hyper-heuristic models that select algorithms based on game characteristics, also exhibit potential but are constrained by the generality of features.

Implications and Future Research Directions

The implications of this research extend into AI-assisted game design, with the framework serving as a tool for developing adaptable AI agents that can assist in real-world gaming scenarios. The exploration of procedural game and content generation indicates potential application in automatic game design, enhancing the creativity and efficiency of game development processes.

For future AI developments, the need for more advanced generalization techniques is highlighted, particularly those that can dynamically adapt to varying game environments and player interactions. This includes better game classification systems for improved algorithm selection and the incorporation of robust opponent modeling in two-player games.

Conclusion

Overall, the GVGAI framework stands as an influential platform in general AI research, providing crucial insights into multi-faceted AI development challenges. Continued research and competition in this space promise to push forward the capabilities of AI in learning, planning, and content generation within the dynamic context of video games.