Skip to content

README.md

Latest commit

 

History

History
44 lines (33 loc) · 3.12 KB

README.md

File metadata and controls

44 lines (33 loc) · 3.12 KB

Machine Learning Path Selector

Entry point

To run tests with this path selector use jarRunner.kt. You can pass a path to a configuration json as the first argument. Gathered statistics will be put in a folder according to your configuration.

Config

A config object is declared inside MLConfig.kt. A detailed description of all the options is listed below:

  • gameEnvPath - a path to a folder that contains trained models (rnn_cell.onnx, gnn_model.onnx, actor_model.onnx) and a blacklist of tests to be skipped (blacklist.txt), also some logs are saved to this folder
  • dataPath - a path to a folder to save all statistics into
  • defaultAlgorithm - an algorithm to use if a trained model is not found, must be one of: BFS, ForkDepthRandom
  • postprocessing - how actor model's outputs should be processed, must be one of: Argmax (choose an id of the maximum value), Softmax (sample from a distribution derived from the outputs via the softmax), None (sample from the outputs — only when they form a distribution)
  • mode - a mode for jarRunner.kt, must be one of: Calculation (to calculate statistics used to train models), Aggregation (to aggregate statistics for different tests into one file), Both (to both calculate statistics and aggregate them), Test (to test this path selector with different time limits and compare it to other path selectors)
  • logFeatures - whether to save statistics used to train models
  • shuffleTests - whether to shuffle tests before running (affects the tests being run if the dataConsumption option is less than 100)
  • discounts - time discounts used when testing path selectors
  • inputShape - an input shape of an actor model
  • maxAttentionLength - a maximum attention length of a PPO actor model
  • useGnn - whether to use a GNN model
  • dataConsumption - a percentage of tests to run
  • hardTimeLimit - a time limit for one test
  • solverTimeLimit - a time limit for one solver call
  • maxConcurrency - a maximum number of threads running different tests concurrently
  • graphUpdate - when to update block graph data, must be one of: Once (at the beginning of a test), TestGeneration (every time a new test is generated)
  • logGraphFeatuers - whether to save graph statistics used to train a GNN model to a dataset file
  • gnnFeaturesCount - a number of features that a GNN model returns
  • useRnn - whether to use an RNN model
  • rnnStateShape - a shape of an RNN state
  • rnnFeaturesCount - a number of features that an RNN model returns
  • inputJars - jars and their packages to run tests on

How to modify the metric

To modify the metric you may change values of the reward property of the ActionData objects. They are written inside the property path of the FeaturesLoggingPathSelector. Currently, the metric is calculated in the remove method of the FeaturesLoggingPathSelector.

Training environment

The training environment and its description are inside environment.zip.

"Modified" files

Source files which names start with "Modified" are modified copies of files from other modules. They were modified to support this path selector.