Skip to content

Features and Limitations

Features

Pause and resume

realm-tune can be paused and resumed. To pause, an interrupt signal can be sent to the process (e.g., ctrl+c in the terminal). The key to enable realm-tune to properly resume is to ensure that the output_path is pointed to the directory where the previous runs were saved. By default, the output directory will have resemblance to the following: <some_path>/runs/RealmTune_xx-xx-xxxx_xx-xx-xx.

The output_path variable can be set through the cli args (i.e., --output-path) or through the config file (see realm-tune's config documentation)

Note: This feature works on Linux and MacOS, but on Windows due to the way they handle interrupts, the process may not pause properly. Therefore, we recommend users to use realm-tune with WSL on Windows.

Automatically running full training with best hyperparameters

Due to the computational cost of performing hyperparameter tuning on non-trivial games, we recommend users to reduce the number of training steps per tuning trial, but instead increase the number of tuning trials so that chances of finding a good hyperparameter space is maximized as much as possible.

As such, realm-tune has the ability to automatically initiate a larger run (i.e., more training steps) once all hyperparameter tuning runs have been completed. In this case, realm-tune will use the hyperparameters from the best tuning trial to perform the full training.

This can be configured through the cli args (e.g., --full-run --full-run-max-steps n where n is the number of training steps for the full run) or through the config file.

Note: The full training run can also be paused and resumed.

Limitations

  • Does not support multiplayer environments (i.e., environments with >1 behaviour(s)). realm-tune automatically checks for this in the code, so the program will error out on multiplayer environments.

  • Does not support in-editor training, must be trained on a build.