Agent-Based Playability Verification for Procedural Content Generation

26 concepts20 papers1 proposal

Procedural Content Generation (PCG) allows for the creation of vast, diverse game worlds, but it faces a fundamental challenge: a generator can easily produce a level that looks correct but is impossible to complete. While a human designer can intuitively spot a missing jump or a blocked doorway, automated systems require a formal feedback loop to ensure functional validity.

Togelius et al. argue that games are a unique domain for computational creativity because they provide a closed-loop environment where creative output can be objectively measured through playability. By using automated agents to "play" through generated content, researchers can verify reachability and solvability, ensuring that the interactive experience remains functional.

The Simulation-Based Feedback Loop

The core mechanism of agent-based evaluation is the integration of a simulated player into the generation pipeline. Instead of relying solely on heuristics or visual patterns, the system treats the game engine as a laboratory. Liu et al. et al. describe this in the context of the General Video Game AI (GVGAI) framework, where autonomous agents act as automated testers to validate the quality and balance of new levels or rules.

This feedback loop serves several distinct roles:

  • Feasibility Filtering: In constructive generation, where levels are built step-by-step, agents can identify traversability issues that simple algorithms might miss. Khalifa et al. et al. note that purely constructive methods often lack this validation, risking the production of "broken" content (Two-step Constructive Approaches for Dungeon Generation).
  • Fitness Steering: In evolutionary algorithms, agent performance serves as a fitness function. Volz et al. demonstrated this by using the CMA-ES algorithm to search the latent space of a Generative Adversarial Network (GAN), "steering" the generator toward levels that are not only visually authentic to Super Mario Bros. but also functionally playable.
  • Constraint Satisfaction: Rather than just a goal to optimize, playability can be a hard constraint. Khalifa et al. et al. utilized pathfinding algorithms like A* as a hard constraint during the training of CESAGAN, ensuring the model only learns from valid spatial layouts.

The Negative Proof Problem

A significant hurdle in this research is what Shaker et al. identify as the "negative proof problem." When an automated agent fails to solve a level, it does not necessarily mean the level is unplayable; it might simply mean the agent is not intelligent or precise enough to find the solution.

This is particularly acute in physics-based games like Cut the Rope or Angry Birds. Stephenson et al. argue that the lack of robust AI agents for complex physics environments is a primary bottleneck for PCG, as generators cannot accurately calibrate difficulty or filter impossible levels without reliable solvers. If an agent is too simple, the system may discard high-quality, challenging content that a skilled human could solve, leading to a "false negative" in content filtering.

Real-Time Verification and Co-Creativity

Beyond fully automated generation, agent-based verification is transforming "mixed-initiative" tools, where humans and AI design together. In these systems, the agent acts as a continuous feasibility checker.

For instance, Delarosa et al. introduced RL Brush, which uses Reinforcement Learning agents to provide real-time nudges to designers. By instantly checking if a human's modification has broken the solvability of a puzzle, the tool allows for a more fluid creative process. Similarly, Khalifa et al. et al. argue in their work on Lode Encoder that real-time automated verification is essential to reduce the frustration of manual playtesting during the design cycle.

Calibrating Experience through Personas

The ultimate goal of agent-based evaluation is often not just playability, but the calibration of player experience. By using "procedural personas"—agents modeled after specific playstyles—designers can predict how different types of players will interact with a level.

Togelius et al. suggest that these modeled agents can act as automated critics, helping to evaluate difficulty and "vibe." This approach is further refined in frameworks like Mech-Elites, where Charity et al. et al. use agents to "illuminate" the space of gameplay mechanics, ensuring that generated tutorials effectively teach specific combinations of actions by verifying they are winnable within fixed time budgets.

Key Foundational Work

Evolving Playable Content for Cut the Rope through a Simulation-Based Approach (2013) — contributed 3 concept elements to this direction.
Computational Game Creativity (2014) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 18 papers
  2. Version 1 · 2026-10-04 · 33 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_agent_based_fitness_evaluation_c70277ad,
  title = {agent-based fitness evaluation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/agent-based-fitness-evaluation/rdr_c70277ad9a2c48b6a97ce877ca17bc3a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_agent_based_fitness_evaluation_c70277ad_v2,
  title = {agent-based fitness evaluation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/agent-based-fitness-evaluation/rdr_c70277ad9a2c48b6a97ce877ca17bc3a/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake