Dataset¶
Expand the training dataset and maintain it with helper tools.
Dataset expansion¶
To enhance the diversity and quality of the training data, the dataset is extended with additional C# methods sourced from popular open-source algorithm repositories:
Dataset tools¶
Two helper scripts under tools/dataset_tools/ automate episode generation
and dataset cleaning. They import the agent package, so run them from that
directory with AIAgent on PYTHONPATH:
generate_episodes.pyCreates multiple training episodes for each method by varying the
StepsToStartparameter, yielding a richer set of examples.clean.pyScans training logs and removes episodes where the baseline strategy (BFS/DFS) completed execution before the neural network could participate (indicated by an “immediate GameOver”). Such episodes are useless for training and are discarded.
Both tools rewrite the dataset file atomically: the new content is written
to a temporary file in the same directory and then renamed over the dataset
(os.replace, see tools/dataset_tools/atomic_io.py), so a crash or
KeyboardInterrupt mid-write leaves the previous dataset intact instead
of a truncated file.
clean.py detects bad episodes from the free-form training log; the log
phrasing it matches (immediate GameOver on <MapName>) is pinned by the
named constants at the top of the module until a structured log source
exists.
Basic usage:
cd tools/dataset_tools
PYTHONPATH=../../AIAgent python3 generate_episodes.py
PYTHONPATH=../../AIAgent python3 clean.py