Game Design Meets Machine Learning

Note: This documentation is still under construction. There will be links that go more in depth.

• Measurement Of Fun

  • Session Length -> The more the player is engaged, the longer the player stays.

  • Map Coverage -> The more the player is engaged, the more the player explores.

  • Variety Of Items Collected -> The more the player is engaged, the more the player collect different items.

  • Amount Of Resources Spent -> The more the player is engaged, the more the player spends on resources.

• Intepreting Local And Global Optima In Game Design

  • Local Optima -> The best solution for anyting related to the current game session.

  • Global Optima -> The best solution for all game sessions.

• What’s Your Goal?

  • Reward Maximization -> Use “measurement of fun” metrics as rewards and combine it with reinforcement learning models.

  • Prediction -> Use regression and classification models.

  • Best Middle Values -> Use clustering models.

• Game Environment Data Is Far More Cleaner Than Real World Data

  • Noise usually comes from overlapping interactions.

  • Your model’s global optimum might be a real global optimum.

  • Game environment states are just a series of physics calculations. Your model may accidentally associate certain things with certain states!

• Model Calculation Speed Vs The Game Engine

  • Per Frame (Physics/Render) -> Model must be fast. Ideally use single datapoints or online models here.

  • Per Interval -> Model calculation time must not exceed the interval. Ideally use mini-batch data here.