General Description

This project developed for "Lifelong Self-Adaptation: Self-Adaptation Meets Lifelong Machine Learning" paper, published at SEAMS@ICSE 2022. In this paper, the proposed archtiecture was validated on two varient domains: Internet-of-Things and Gas Delivery System. You can find these two validation cases in two different folders in src. Moreover, extra materials (required data for IoT case) can be found on the webiste of the paper.

IoT Validation Case

Porject Structure

In the root folder, there are multiple folders:

1. deltaiot_simulator is the DeltaIoT simulator, the IoT system for validating the lifelong learning loop.
2. self_adaptation implements MAPE-K functions (feedback_loop.py) with support of machine learning (machine learning component is implemented in machine_learning.py).
3. lifelong_self_adaptation implements a meta layer over the self-adaptation for the matter of lifelong self-adaptation. It comprisies:
   1. knowledge component (knowledge.py),
   2. task manager component (task_manager.py),
   3. knowledge-based reasoner (a function in lifelong_learning_loop.py)
   4. task-specific miner (a function in lifelong_learning_loop.py)
   5. general lifelong learning loop (in lifelong_learning_loop.py)
4. common implements base classes that are shared between the adaptation layers. For example, goal types (goal.py) and learner types (learning.py). The other important part of this module is stream that mimics collecting data in each adaptation cycle.
5. visualizer implements some methods to visualize the adaptation result

Requirements

This project is developed based on the key following tools and platforms:

1. Ubuntu 20.04 LTS
2. Python 3.7
   1. Plotly library for data visualization
   2. Orca enging to render PDF files (https://community.plotly.com/t/using-orca-to-export-plotly-plots-to-pdf-files/18816)

Also, we used many other standard libraries in python like numpy, but there should not be a complexity for getting final resutls.

Run

To start the project, first, you need to run the DeltaIoT system simulator (in deltaiot_simulator folder) based on

• some desired configs in deltaiot_simulator > SMCConfig.properties file. (The current config shows DeltaIoTv1 simulator with 1505 cycles.)
• Another required modification before executing the simulator is changing the global interference and their affect over links in deltaiot_simulator > activforms > src > mapek > FeedbackLoo.java file. You can switch between no drift and under drift scenarios by setting the variable is_drift_scenario in the function start() of the same folder.
• To collect all quality services for all adaptation option, Uppaal-SMC has been utilized. To run this verifer, it is rquired to address verifyta file which provides the way of running the verifier through the command line. This file has been already located at deltaiot_simulator > uppaal-verifyta > verifyta. But, depending on the platofrms and their correspnding settings, you may need to refer into the file by an absolute addressing. To do that, you can place the absolute address of this file in deltaiot_simulator > activforms > src > smc > SMCChecker and at the starting point of "./uppaal-verifyta", "verifyta -a %f -E %f -u %s".
• Eventually, for running the simulator, you need to run the Main.java class in deltaiot_simulator > activforms > src > main folder. (If you install the VS code IDE and the Extension Pack for Java for it, by opening the project from the root of deltaiot_simulator, you can simply right click over the file name and click on Run Java to run the project.) You can also use mvn exec:java -pl activforms command to run the simulator by maven. More details of running the simulator can be found here https://people.cs.kuleuven.be/~danny.weyns/software/machine-learning.htm. You can have access to pre-generated data for both above-mentioned cases next to this project zip file on the same website. (It is also a zip file contains two folders for the two different scenarios.)

Second, after running the simulator, some files will be generated in deltaiot_simulator > output > deltaiotv1 (if the setup of the simulator in SMCConfig.properties is set on DeltaIoTv1). These files contains true (or simulated) values of qualities of the system for each adaptation option and value of uncertainties (links' interference and motes' load of messages) in each adaptation cycle. We should move them into a subfolder of the ./data folder to be used for further feedback loop simulations. Also, regarding to the paper, we need to run the first step twice, one for a no drift scenario and other for under derfit scenario. Hence, we should move the generated files each time, as they can be rewritten by the next running simulation.

Thrid, after moving required data, we should the address of files in fetch_stream() function in ./main.py. There are two scenarios inside this function that can be adjusted by the input parameter with_drift to it. After setting required addresses in this function, we are ready to run a self-adaptation loop with/without meta-layer that is for lifelong self-adaptation. For this purpose, there are two functions in the ./main.py called run_feedbackloo and run_ll_loop, respectively. Before calling these functions in the ./main.py, you should set the address for their output files which will contain their predictions for quality of services for each adaptation option in each adaptation cycle. For running the desired loop, you only need to uncomment the corresponding function to the loop. For example, for running MAPE-K feedback loop without meta-layer lifelong learning, you need to call run_feedbackloop() function in the scope of if __name__ == "__main__":. Note that you need to adjust the address of the cycle data in the corresponding Stream object of the adaptation function. As another example, you can run the self-adaptation loo with lifelong self-adaptation meta-layer using run_ll_loop() function. Note that, you can also config more settings for the loops (MAPE-K and lifelong learning) via input parameters of their corresponding object. For example, for a lifelong learning loop, you can determine that the loop must evovle the MAPE-K loop by selecting the best function among one or two types of learner, using bilearning_mode = False and bilearning_mode = True, respectively. Running of the ./main.py can be done by the following command based on the the root folder:

$ python3 main.py

Eventually, you can compare and visualize the output of loops' using qos_visualization_multiple_goals_with_utility() function in main.py. Note that the corresponding file addresses to prediction outputsfor different scenarios should be adjusted beforehand (in appropriate placeholders determined by <path to the generated file>).

These are main steps for simulating and comparing results of different scenarios under different type of loops (i.e., MAPE-K loop or the loop under lifelong learning with different settings). Also, there other functions to visulize training time in these scenarios and also representing global network interference (affects on SNRs) and task identification in different adaptation cycles by snr_vis(), plot_learning_time(), and plot_drifts() functions, respectively. Note that you should update address of the related generated files for each of these functions based on the exiting placeholders in those functions.

Results

For details of plots in the paper, you can look at the results in the figures folder. You can find more interactive figures in .htm formats in compare with .pdf versions. Also, name of these files are started with the Fig#i which #i indicates the figure number of plots in the paper or its relation to #i-th figure number. For example, Fig5-energy_consumption has not been included in the paper, but it is showing the same meaning of the figure 5 for the packet loss in the paper.

Gas Delivery System Validation Case

Porject Structure

In the root folder, there are multiple folders:

1. self_adaptation implements MAPE-K functions (feedback_loop.py) with support of machine learning (a classifier to predict class of sensory data).
2. lifelong_self_adaptation implements a meta layer over the self-adaptation for the matter of lifelong self-adaptation. It comprisies:
   1. knowledge component (knowledge.py),
   2. task manager component (task_manager.py),
   3. knowledge-based reasoner (a function in lifelong_learning_loop.py)
   4. task-specific miner (a function in lifelong_learning_loop.py)
   5. general lifelong learning loop (in lifelong_learning_loop.py)
3. common implements base classes that are shared between the adaptation layers. The important part of this module is stream that mimics collecting data in each adaptation cycle (data from this link: http://archive.ics.uci.edu/ml/datasets/Gas+Sensor+Array+Drift+Dataset+at+Different+Concentrations)
4. visualizer implements some methods to visualize the adaptation results in the paper.

Requirements

This project is developed based on the key following tools and platforms:

1. Ubuntu 20.04 LTS
2. Python 3.7
   1. Plotly library for data visualization
   2. Orca enging to render PDF files (https://community.plotly.com/t/using-orca-to-export-plotly-plots-to-pdf-files/18816)

Also, we used many other standard libraries in python like numpy and keras_tuner, but there should not be a complexity for getting final resutls.

Run

To start the project, you onlu need to run the following command from the root folder python3 main.py.

Results

For details of plots in the paper, you can look at the results in the figures folder. You can find more interactive figures in .htm formats in compare with .pdf versions. Also, name of these files are started with the Fig#i which #i indicates the figure number of plots in the paper or its relation to #i-th figure number. For example, Fig13_accuracy_over_time.pdf has not been included in the paper, but it is showing the accuracy of different methods over time which their values are represented in different ranges in Fig13.

Citation

@inproceedings{gheibi2022lifelong,
    title={Lifelong Self-Adaptation: Self-Adaptation Meets Lifelong Machine Learning},
    booktitle={2022 International Symposium on Software Engineering for Adaptive and Self-Managing Systems (SEAMS)},
    year={2022},
    organization={ACM}
}