Skip to content

Latest commit

 

History

History
103 lines (65 loc) · 4.54 KB

README.md

File metadata and controls

103 lines (65 loc) · 4.54 KB

Data Generation

Notes on dataGen.js.

Purpose

dataGen.js provides any functionality needed to generate data related to the proxy coaching experiments presented in the corresponding paper.

Dependencies

In order to run the dataGen.js script, you need to have installed npm - instructions here. Furthermore, the script depends on the following scripts to run properly:

  • fs, which is a native node.js script;
  • perf-hooks, which is a native node.js script;
  • parsers.j, which is a Prudens-related script, found in prudens-local/parsers.js;
  • prudens.j, which is a Prudens-related script, found in prudens-local/prudens.js;
  • decision-tree, which may be found here.

Provided that the repository's directory structure is kept intact, there is no further action needed to be taken so as to run dataGen.js.

Execution

Policy Generation

In order to generate policies according to the described protocol presented in the paper, you may cd to the directory where dataGen.js is located and run the following command:

npx run-func dataGen.js stackGen

If promted to install run-func, press y (confirm). Any results will be written at data/policies/genKbs.json.

Remark: Since there are some random choices made during the data generation process, the resulting policies might not be identical to the ones used in the original experiments. The original datasets and any related metadata used may be found at /used-data.

Train-Test Dataset Generation

In order to generate train and test datasets from a set of policies, run the following command:

npx run-func dataGen.js generateTrainTest

To facilitate reproducibility, by default generateTrainTest uses the originally used data found at /used-data. However, in case you wish to use other policy sets, you may pass the corresponding paths as parameters, as follows:

npx run-func dataGen.js generateTrainTest "policyTagsPath" "policiesPath"

In the above policiesPath is a string corresponding to the path of policies while policyTagsPath is a string corresponding to the path of a file containing the tags of the policies to be used - as in /used-data/kbs.json.

Any results will be written to data/datasets/trainTest.json.

Coaching Dataset Generation

In order to generate the coaching dataset from a set of policies, run the following command:

npx run-func dataGen.js generateCoaching

To facilitate reproducibility, by default generateCoaching uses the originally used data found at /used-data. However, in case you wish to use other policy sets, you may pass the corresponding paths as parameters, as follows:

npx run-func dataGen.js generateTrainTest "n" "policyTagsPath" "policiesPath"

In the above, n is the number of coaching sets to be generated, policiesPath is a string corresponding to the path of policies while policyTagsPath is a string corresponding to the path of a file containing the tags of the policies to be used - as in /used-data/kbs.json.

Any results will be written to data/datasets/coaching.json.

Proxy Coach Generation

Tree Training

To train trees as proxy coaches, run the following command:

npx run-func dataGen.js trainTrees

As with generateTrainTest, trainTrees by default uses data used in the original experiments, to facilitate reproducibility. In case you wish to run the experiments with your own data, provide the corresponding paths as parameters, as shown below:

npx run-func dataGen.js trainTrees "policyTagsPath" "policiesPath" "trainTestPath"

policyTagsPath and policiesPath are as above, while trainTestPath corresponds to the path of a file containing labelled train and test sets, as, e.g., /used-data/trainTest.json.

Any results will be written at data/coaches/trainedTrees.json.

Forest Training

To train random forests as proxy coaches, run the following command:

npx run-func dataGen.js trainForests

As with generateTrainTest, trainForests by default uses data used in the original experiments, to facilitate reproducibility. In case you wish to run the experiments with your own data, provide the corresponding paths as parameters, as shown below:

npx run-func dataGen.js trainForests "policyTagsPath" "policiesPath" "trainTestPath"

policyTagsPath and policiesPath are as above, while trainTestPath corresponds to the path of a file containing labelled train and test sets, as, e.g., /used-data/trainTest.json.

Any results will be written at data/coaches/trainedForests.json.