This project shows the prototype MoConfig. MoConfigSampling (in Java) and MoConfigComparison (in Python) are two main parts, implementing the process of MoConfig sampling and the comparasion with the rank-based method, respectively.
The 20 raw datasets we used are saved in the directory MoConfigComparison/raw_data/.
The environments of MoConfig project includes JDK 1.7/1.8, Python 3.X. Moreover, sklearn, pandas, numpy libraries should also be installed.
Step 1: We run the rank-based method to divide each dataset into the training pool, validation pool, and test pool. Above sets are generated and saved in the directory MoConfigComparison/parse_data/.
Step 2: We combine the training pool and validation pool generated by rank-based method into a combiantion dataset (MoConfigComparison/parse_data/combined_train/). We copy this combination dataset to the directory MoConfigSampling/testbed/input/ and take it as the input of MoConfig.
Step 3: We run the MoConfig to sample the optimal sampling sets from the training pool. The optimal sampling sets are saved in the directory MoConfigSampling/testbed/output/.
Step 4: We find the best configuration based on the sampling sets generated by MoConfig. The prediction results are saved in MoConfigComparison/experiments/moconfig/.
Step 5: We compare the results between the rank-based and MoConfig methods. The visualized results of comparision are saved in MoConfigComparison/pics/.
Here, we list the related code entries to above steps. For instance, we just run the main function in rankbased.py to implement the Step 1 and Step 2.
| Step | Running entry | Time cost |
|---|---|---|
| 1,2 | MoConfigComparison/rankbased.py |
Around 150 seconds |
| 3 | MoConfigSampling/src/main/java/cn/edu/whu/cstar/experiments/Launcher.java |
Around 20 minutes |
| 4 | MoConfigComparison/moconfig.py |
Around 40 minutes |
| 5 | MoConfigComparison/experiments.py |
Arounds 150 seconds |