Anfan

Reinforcement learning Video game AI

Agents that learn by playing.

Anfan studies how interactive game agents learn from observations, actions, and feedback—and how to tell whether that learning holds up beyond one level, one seed, or one score.

Read the research agenda Discuss research
The agent loop Observe. Act. Learn. Repeat.
  1. 01
    Observestate, pixels, events
  2. 02
    Decidepolicy selects an action
  3. 03
    Interactthe game changes
  4. 04
    Updatefeedback shapes learning

The next state starts the loop again.

01 Research

Video games make hard learning problems playable.

Games combine sequential decisions, partial information, changing goals, and immediate interaction. They are useful environments for asking concrete questions about reinforcement learning.

  1. Learning over long horizons

    How can an agent connect an action now with an outcome much later, especially when rewards are sparse or delayed?

  2. Generalizing beyond a run

    Does a policy still work when levels, opponents, starting states, or game conditions change?

  3. Adapting while interacting

    Can an agent respond to new situations during play instead of relying only on behavior memorized during training?

  4. Evaluation that explains failure

    Scores matter, but so do repeated trials, held-out scenarios, learning efficiency, behavior traces, and clear accounts of where an agent breaks.

02 Process

From game environment to inspectable evidence.

The aim is not a polished demo with an unclear training story. Each study should make the learning problem, comparison, and evidence legible.

Environment

Define the interaction

Specify observations, actions, rewards, episode boundaries, seeds, and the conditions used for training and evaluation.

Learning

Establish useful baselines

Compare reinforcement-learning policies against simple, reproducible baselines before adding complexity.

Evaluation

Test more than the best score

Use repeated runs and held-out conditions, then report variability, sample use, failure modes, and relevant compute.

Release

Connect claims to artifacts

When work is ready, pair the result with the code, configuration, checkpoints, and traces needed to inspect it.

03 Projects

Selected projects

A few examples of our work in reinforcement learning and interactive game systems.

Combat Learning Agent

A UE5 reinforcement-learning agent trained with PPO for real-time melee combat.

View project

Decision Scheduling for RL Game Agents

Research on when real-time RL agents should make their next decision as actions and interruptions unfold.

View project

Motion-Controlled Medieval Combat

An Unreal Engine C++ prototype using gyroscope input for multiplayer melee combat.

View project

04 Contact

Talk about game agents.

For research conversations, environments, evaluation, or future collaboration, email Anfan directly.

Research research@anfan.org
General hello@anfan.org