SEPIA: A Scalable Game Environment for Artificial Intelligence Teaching and Research

AAAI 2013

TL;DR

This paper introduces SEPIA, a novel game environment built on real-time strategy (RTS) games, designed specifically for teaching and researching Artificial Intelligence (AI). SEPIA offers a flexible and scalable platform that allows for the creation of diverse AI challenges, from introductory undergraduate assignments to complex graduate-level research problems, bridging the gap between engaging gameplay and rigorous AI development.

What Problem Does This Paper Address?

Traditional AI teaching often relies on simplified, abstract problems (like the 8-puzzle) that may not fully capture the complexity of real-world decision-making or engage students as effectively. While more complex games like Pac-Man and Mario have been explored, they may not offer the depth and breadth of challenges required for advanced AI research. The authors identified a need for a platform that is:

  • Pedagogically Scalable: Adaptable to various skill levels, from beginners to advanced researchers.
  • Engaging: Leverages the inherent appeal of video games to motivate students.
  • Research-Oriented: Provides the necessary features for developing and testing sophisticated AI agents, particularly in dynamic, multi-agent environments.
  • Flexible: Allows for easy modification of game mechanics and scenarios to create tailored learning experiences.

Existing RTS engines and APIs were found to be either too focused on human play, lacking essential simulation features for AI, or too proprietary and difficult to integrate with.

Methods

SEPIA is a Java-based game environment built around the core mechanics of Real-Time Strategy (RTS) games. Key features include:

  • RTS Core Mechanics: Incorporates resource gathering, unit and building production, combat, and technology research, creating a complex interplay of economic, combat, and research sub-games.
  • Configurability: Uses XML configuration files to define maps, player teams, agent classes, victory conditions, and game parameters (e.g., time limits).
  • Scalable Difficulty: Game mechanics can be simplified (e.g., disabling partial observability, technology trees, or specific units) to adjust difficulty for different learning levels.
  • Agent Interface: A standard Agent interface allows students to write their own AI agents in Java without needing to recompile the entire SEPIA engine.
  • Simulation Support: Provides robust simulation capabilities, allowing agents to internally model the environment, test actions, and observe outcomes without affecting the real game state. This is crucial for techniques like Monte Carlo planning.
  • Execution Strategies: Supports turn-based (round-robin) and simultaneous action execution, with mechanisms to ensure state consistency.
  • Graphical Interface (VisualAgent): An optional visual agent that can be attached to any AI agent, allowing for debugging and visualization of the game state from the agent's perspective.
  • Logging and History: Detailed logging and save functionality (game state, agent state, history) are provided to aid debugging and enable techniques like action replay.
  • Compound Actions: Agents can specify compound actions, which SEPIA's internal planner resolves into primitive actions.

Datasets

Not discussed in the paper. The paper focuses on the creation of scenarios within the SEPIA environment rather than using pre-existing datasets for training or evaluation.

Key Insights and Ideas

  • "Pedagogically Scalable" Domains are Key: The authors emphasize that by slightly altering the mechanics of a complex domain like RTS, one can create a spectrum of problems suitable for different AI learning levels. This is more effective than switching between entirely different domains.
  • Bridging the "Coolness" Gap: Using video game environments like RTS games makes AI concepts more relatable and engaging for modern students, who are often already familiar with and enjoy playing such games.
  • RTS as an AI Testbed: The dynamic, multi-agent, and partially observable nature of RTS games presents significant challenges that are representative of many hard AI problems, making it an excellent domain for both teaching and research.
  • Custom Environment Benefits: Developing a custom environment like SEPIA, tailored specifically for AI development, offers greater flexibility and control over simulation features, agent interfaces, and experimental infrastructure compared to adapting existing game engines.

Key Results

The paper details the use of SEPIA in undergraduate and graduate AI courses.

  • Undergraduate Course: Five programming assignments were designed, progressively covering search (A*), game trees (alpha-beta pruning), automated planning, probabilistic inference (POMDPs), and reinforcement learning (Q-learning). Student feedback was overwhelmingly positive, with many praising the assignments as enjoyable and effective learning tools. The course rating saw a slight improvement after SEPIA's introduction.
  • Graduate Course: SEPIA was used for assignments on advanced planning (online planning for resource production) and reinforcement learning (policy gradient for combat). The response was more mixed, partly due to some students lacking the prerequisite undergraduate AI course and thus being unfamiliar with SEPIA. The authors plan to address this with tutorials.

Limitations and Future Work

  • Communication: SEPIA currently does not explicitly support communication between competing agents, limiting its use for teaching concepts like diplomacy or negotiation.
  • Documentation and API: Initial versions had documentation gaps and missing API functionality (e.g., object copying), which caused student frustration. These have since been addressed.
  • Graduate Student Familiarity: The mixed response in the graduate course highlights the need for better onboarding for students who haven't used SEPIA before.
  • Future Work: The authors plan to continue developing SEPIA by adding more features and interesting scenarios to demonstrate AI algorithms. They also aim to make it available to other instructors.

Summary updated

BibTeX
@inproceedings{2013sepia,
  title = {{SEPIA: A Scalable Game Environment for Artificial Intelligence Teaching and Research}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2013},
}
Figure