A Monte Carlo Tree Search conceptual framework for feature model analyses

This repository contains all the resources and artifacts of the Monte Carlo Tree Search conceptual framework, supporting the papers entitled A Monte Carlo Tree Search Conceptual Framework for Feature Model Analyses submmited to the Journal of Systems and Software (Special Issue on Systems and Software Product Lines of the Future); and the paper Monte Carlo Tree Search for feature model analyses: a new framework for decision-making published in the 25th International Systems and Software Product Line Conference (SPLC 2021) by the authors José Miguel Horcas, José A. Galindo, Ruben Heradio, David Fernández-Amoros, and David Benavides.

Links to published material:

JSS journal paper: https://doi.org/10.1016/j.jss.2022.111551
SPLC 2021 best paper awarded paper: https://doi.org/10.1145/3461001.3471146
Video with the presentation: https://www.youtube.com/watch?v=a7kYYQHbnZo

We present a Monte Carlo conceptual framework to analyze SPL problems by modeling them as a sequence of decision steps and solving them with Monte Carlo techniques. The Monte Carlo framework has been developed on top of the Python framework for automated analysis of feature models proposed by Galindo and Benavides. The details of the core components of the Monte Carlo framework are described here.

This repository is organized into three parts:

The Monte Carlo conceptual framework. This includes:

A set of interfaces to be implemented in order to model SPL problems as sequences of (state, actions) pairs.
An implementation of several Monte Carlo methods, including the Monte Carlo Tree Search (MCTS) method, ready to be used to solve any problem that implements the aforementioned interfaces.

Instantiations of the framework to analyze different problems. In particular, we provide two concrete implementations of the (state, actions) interfaces:

An implementation of (state, actions), where states represent configurations of a feature model.
An implementation of (state, actions), where states represent feature models.

Results of the analyses performed with Monte Carlo techniques over the problems modeled.

The Monte Carlo conceptual framework

Interfaces

The framework defines the three main interfaces (State, Action, and Problem) to be implemented in order to model and solve a problem with Monte Carlo methods. The State interface specifies the necessary methods to explore the search space so that from a given initial state, we can reach all states. The State interface has to be implemented only once by defining the state transition function (find_successors() and find_random_successor()), the is_terminal() condition, and the reward() function. The Action interface is defined for each applicable action. The Problem interface specifies how a problem is setup at the start by providing an initial state, and how a solution is decoded from a terminal state to represent the solution in a human-readable form.

Algorithms and Monte Carlo methods

Although the framework focuses on Monte Carlo methods, it can support any search-based algorithm that is built using the previous interfaces, such as a classical A-star search (A*) or genetic algorithms. The Algorithm, *Monte Carlo and Monte Carlo Tree Search interfaces provide a base abstract implementations of any search-based algorithm, Monte Carlo, and MCTS method, respectively, which can be specialized with different algorithms and variants of Monte Carlo methods. The following generic search-based algorithms are currently available:

Random strategy: It chooses a random action from the current state without running any simulation. This is not a Monte Carlo method itself, but it is often used in game theory to simulate a random player and it is widely used as a baseline to compare Monte Carlo methods.
A-star search: It is an implementation of the most widely known form of best-first search. It evaluates nodes by combining the cost to reach the current node and a heuristic of the cost to get from the current node to a terminal node. This method is available only for testing the correctness of the implementation of the states and the actions because this method performs an exhaustive search, which makes it infeasible in practice for medium and large search spaces.

An implementation of several Monte Carlo methods, including MCTS, is available to solve any problem that implements the aforementioned interfaces. The Monte Carlo and MCTS methods available in the framework are the following:

UCT Algorithm: An implementation of MCTS that builds a search tree and uses the upper confidence bound for trees (UCT) selection strategy. This strategy favors actions with a higher Q-value but allows at the same time to explore those actions that have not yet been sufficiently explored.
Greedy MCTS: A best-first strategy that favors exploitation against exploration.
flat Monte Carlo: A basic Monte Carlo method with random action selection and no tree growth.

Configuration

Each search-based algorithm, including the Monte Carlo methods, can be configured with a stopping condition that specifies a computational budget (e.g., time, memory, or a number of steps) to find a solution, after which the algorithm will return the solution (if found). If not stopping condition is provided the algorithm will run until a solution is found. In addition, each Monte Carlo method can be configured with a second stopping condition that also allows specifying a computational budget (e.g., number of simulations or time to make a decision during each algorithm step), so that when the condition is reached, the Monte Carlo method will return the best decision found. Currently, two different stopping conditions are implemented:

An Iterations constraint which allows specifying a maximum number of steps for the algorithm to find a solution or a maximum number of simulations or iterations for Monte Carlo methods to make a decision.
An Time constraint that allows specifying a maximum execution time in seconds for the whole search or for making a decision in Monte Carlo methods.

Moreover, each Monte Carlo method can also be configured with a selection criterion for the best action decision. For example, select the child with the highest reward, the most visited child, the child with both the highest visit count and the highest reward, or the child that maximizes a lower confidence bound. Currently, the selection criteria that returns the child with the highest reward is available.

Finally, additional configuration parameters can be provided to Monte Carlo methods. Concretely, due to Monte Carlo methods rely on randomness, we provide a seed parameter to initialize the random generator and enable experimental reproducibility. Also, a number of runs can be specified (default is 1) to repeat an experiment several times (e.g., 30) and obtain the median, means and standard deviation of different statistics, as explained below.

The details of the core components of the framework are described here.

Knowledge inference

Two main usages of the Monte Carlo methods are available. First, Monte Carlo methods can be used as a search-based algorithm so that from a given initial state, the algorithm will look for the best solution(s), i.e., terminal state(s). Given an initial state, the algorithm will run until some predefined computational budget is reached or until a terminal state is reached. In each step, the algorithm calls the MCTS method in charge of choosing the best action. In addition to the best solutions found by the algorithm, the MCTS framework provides further knowledge information gathered during the search. The partially optimal decisions made step by step are available, so that we can observe, for example, which feature has been selected in each step during the configuration process for configuration-based analyses, or which relation has been added to the feature model in each step for evolution-based problems or reverse engineering problems.

Second, Monte Carlo methods can be also used to analyze a particular state and its possible alternatives. Given a state, the method will analyze the possible alternatives that can be reached from this state and will return information about the best choices, so that the practitioner can make better decisions. At this respect, the MCTS framework also reports the information stored in the tree search about the Q-values and visit accounts of each (state, action) pair gathered by the MCTS method. To illustrate the knowledge stored in the tree search, we use a data visualization technique called heatmaps, which encodes quantitative values as colors (like in weather maps), so that it compacts large amounts of information (the Q-values) to bring out coherent patterns in the data (e.g., optimal feature selection over the feature model).

Finally, the MCTS framework also reports statistical information about the execution time and memory consumption of the algorithms, such as the median, mean, and standard deviation for both the global solution and each step-wise decision, that allows a deep study of the different Monte Carlo methods and analysis problems.

Installation and execution

Requirements

The implementation of this conceptual framework has relied on the Python programming language. By convention, all requirements are depicted in the requirements.txt file. In particular, the dependencies are:

The framework has been tested in Linux and Windows

Download and setup

To use the framework under Python 3.9, you will need the following:

git clone https://github.com/diverso-lab/fm_montecarlo.git
cd fm_montecarlo

At this point, it is recommended to use a virtual environment:

Create the virtual environment: python -m venv env

Activate the environment (Linux): source env/bin/activate

Activate the environment (Windows): .\env\Scripts\Activate

Finally, install the dependencies: pip install -r requirements.txt

Analyzing problems with the Monte Carlo framework

The following use case diagram shows the four problems that have been implemented, and the output results obtained from applying Monte Carlo methods.

Configuration based analyses
- Localizing defective configurations: This problem consists in identifying the feature model configurations that lead to a given defect or to some other undesired program behaviors. Those defects may happen due to incompatibilities of features, anomalies or errors when the configuration is compiled, deployed or executed. Two real-world feature models are analyzed: the jHipster and the Python framework for AAFMs.
  
  To analyze the feature model of the Python framework for AAFMs, execute:
  
  python main_localizing_defective_configs.py
  
  To analyze the jHipster feature model, run:
  
  python main_jhipster_localizing_defective_configs.py
  
  The analysis can be configured with the following parameters:
  
  -it ITERATIONS: specifies the number of simulations to be executed by the Monte Carlo method (default 100).
  
  -ew EXPLORATION_WEIGHT: the exploration weight constant for MCTS to balance exploitation vs. exploration (default 0.5).
  
  -m METHOD: the Monte Carlo method to be executed: "MCTS" for the UCT Algorithm (default), "Greedy" for the Greedy MCTS, and "flat" for the basic Monte Carlo method.
  
  Additionally, the case study of the Python framework for AAFMs can be configured to use the complete version of the feature model (default) or the excerpt version presented in the paper (using the -e option).
  - Completion of partial configurations: The problem of completing partial configurations deals with finding the set of non-selected features necessary for getting a complete valid configuration. While in a complete configurationk, each feature is decided to be either present or absent, in partial configurations, some features are undecided. So, given a feature model and a partial configuration, we can use Monte Carlo methods to complete the given partial configuration with valid selections.
  The problem can be executed with:
  
  python main_completion_partial_configs.py -fm feature_model -cnf cnf_model -f features
  
  The feature_model parameter is mandatory and specifies the file path of the feature model in FeatureIDE format.
  
  The cnf_model is optional and specifies the feature model in CNF with FeatureIDE (textual) format. This parameter is only required if the feature model has complex constraints (others than "requires" and "excludes").
  
  The features parameter is optional. It is a list of the user's feature selection that represents the initial partial configuration. If it is not provided, the empty configuration is used by default.
  
  The analysis can also be configured with the following parameters:
  
  -it ITERATIONS: specifies the number of simulations to be executed by the Monte Carlo method (default 100).
  
  -ew EXPLORATION_WEIGHT: the exploration weight constant for MCTS to balance exploitation vs. exploration (default 0.5).
  
  -m METHOD: the Monte Carlo method to be executed: "MCTS" for the UCT Algorithm (default), "Greedy" for the Greedy MCTS, and "flat" for the basic Monte Carlo method.
- Minimizing valid configurations: This problem consists in finding a valid configuration with the minimum number of features.
  
  This problem can be executed as the previous one to complete partial configurations, but using the -min option to indicate that the number of feature selections must be minimized:
  
  python main_completion_partial_configs.py -fm feature_model -cnf cnf_model -f features -min
Feature models based analysis
- Reverse engineering of feature models: A well-known problem in SPLs is to synthesize a feature model from a set of configurations automatically. Given a set of feature combinations present in a SPL (i.e., a set of configurations), the goal is to extract a feature model representing all the configurations. The problem can be executed with:
  
  python main_reverse_engineering_fms.py -fm feature_model -cnf cnf_model
  
  The feature_model parameter is mandatory and specifies the file path of the feature model in FeatureIDE format.
  
  The cnf_model is optional and specifies the feature model in CNF with FeatureIDE (textual) format. This parameter is only required when the feature model has complex constraints (others than "requires" and "excludes").
  
  We use all configurations of the given feature model as input configurations to extract a new feature model.
  
  The analysis can also be configured with the following parameters:
  
  -it ITERATIONS: specifies the number of simulations to be executed by the Monte Carlo method (default 100).
  
  -ew EXPLORATION_WEIGHT: the exploration weight constant for MCTS to balance exploitation vs. exploration (default 0.5).
  
  -m METHOD: the Monte Carlo method to be executed: "MCTS" for the UCT Algorithm (default), "Greedy" for the Greedy MCTS, and "flat" for the basic Monte Carlo method.

Results

The analyses provide four kinds of results:

The optimal solution and the partially optimal decisions made step by step are shown in the terminal.

This result is shown for all types of analyses.

Two .csv files with statistics of the analysis are generated in the output_results/stats folder.

This result is also generated for all types of analyses.

A set of heat maps are also generated containing valuable information about each decision made in each step. A heat map file in .csv is generated for each step in the output_results/heatmaps folder. The heat maps contain the normalized Q-value for each decision and a mapping to a warm-cold colors scale which can be used to color a feature model.

This result is only generated for the configuration-based analyses using the Monte Carlo Tree Search method or one of its variants that builds a search-tree (i.e., the heatmaps are not generated for flat Monte Carlo).

In the problem of the reverse engineering feature models, the output is a feature model automatically extracted (in the new UVL format); and a .log file with all the decisions taken during the process, as well as all alternative decisions considered with their normalized Q-values.

Experiment replication

The results from the paper are available in the Results folder. It is worthy to highlight that there are two types of experiments: (1) those about the problem analyses previously presented, and (2) experiments to compare Monte Carlo methods used in the evaluation. Here we explain how to replicate both experiments using the Python scripts provided.

In order to allow replicating the results, we have added an optional parameter -s SEED in all experiments (analyses and comparison) to set up the random seed.

-s SEED: the seed to be used for the random module to replicate the experiments (default None).

To obtain the same results, use the seed 2021 in all cases.

Note: despite setting up the random seed used by the Monte Carlo methods, some results may present slight variations due to the inherent randomness nature of the Monte Carlo methods (see note about the randomness of Monte Carlo methods at the end of this file for more details).

Replicating results from problem analyses

To replicate these experiments, execute the analyses with the following parameters:

Localizing defective configurations (results here):
- For the AAFMs Python Framework feature model (excerpt version):
  
  python main_localizing_defective_configs.py -e -it 100 -s 2021 -m MCTS
Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
- For the AAFMs Python Framework feature model (complete version), use the same command but without the -e parameter:
  
  python main_localizing_defective_configs.py -it 100 -s 2021 -m MCTS
Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
- For the jHipster feature model:
  
  python main_jhipster_localizing_defective_configs.py -it 100 -s 2021 -m MCTS
Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
Completion of partial configurations (results here):
- For the AAFMs Python Framework feature model (excerpt version):
  
  python main_completion_partial_configs.py -fm evaluation/aafmsPythonFramework/model_simple_paper_excerpt.xml -cnf evaluation/aafmsPythonFramework/model_simple_paper_excerpt-cnf.txt -it 100 -s 2021 -m MCTS
  
  Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
  - For the AAFMs Python Framework feature model (complete version):
    
    python main_completion_partial_configs.py -fm evaluation/aafmsPythonFramework/model_paper.xml -cnf evaluation/aafmsPythonFramework/model_paper-cnf.txt -it 100 -s 2021 -m MCTS
  Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
  - For the jHipster feature model:
    
    python main_completion_partial_configs.py -fm evaluation/jhipster/fm-3.6.1refined.xml -cnf evaluation/jhipster/fm-3.6.1refined-cnf.txt -it 100 -s 2021 -m MCTS
  Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
  - Minimizing valid configurations (results here): The commands are the same as the previous problem but using the option -min activated.
    - For the AAFMs Python Framework feature model (excerpt version):
      
      python main_completion_partial_configs.py -fm evaluation/aafmsPythonFramework/model_simple_paper_excerpt.xml -cnf evaluation/aafmsPythonFramework/model_simple_paper_excerpt-cnf.txt -it 100 -s 2021 -min -m MCTS
      
      Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
      - For the AAFMs Python Framework feature model (complete version):
      python main_completion_partial_configs.py -fm evaluation/aafmsPythonFramework/model_paper.xml -cnf evaluation/aafmsPythonFramework/model_paper-cnf.txt -it 100 -s 2021 -min -m MCTS
      
      Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
      - For the jHipster feature model:
      python main_completion_partial_configs.py -fm evaluation/jhipster/fm-3.6.1refined.xml -cnf evaluation/jhipster/fm-3.6.1refined-cnf.txt -it 100 -s 2021 -min -m MCTS
      
      Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
  - Reverse engineering of feature models (results here): For this problem, due to the size of the problem, only the excerpt version of the AAFMs Python Framework is used. Even though, as explained in the paper, this can take a while (around 10-15 min) for 1000 iterations because 1000 feature models are generated in each decision, and all configurations of those feature models are also generated to evaluate the fitness function.
    
    python main_reverse_engineering_fms.py -fm evaluation/aafmsPythonFramework/model_simple_paper_excerpt.xml -cnf evaluation/aafmsPythonFramework/model_simple_paper_excerpt-cnf.txt -it 1000 -s 2021 -m MCTS
    
    Change the -m parameter to flat for the flat Monte Carlo method, and Greedy for the Greedy MCTS.
  Replicating results from the comparison of Monte Carlo methods
  
  The evaluation experiment compares the different Monte Carlo methods over the problem of finding defective configurations in both the AAFMs Python Framework and the jHipster feature models. Results are available here. To replicate the experiments of the evaluation, we provide the following scripts that can be executed as follows:
  - For the AAFMs Python Framework feature model):
    
    python main_comparison_aafm.py -it 1000 -s 2021 -m MCTS
    
    Note that the -it ITERATIONS parameter is the maximum number of iterations/simulations to be performed (5000 default). The script will run the Monte Carlo methods from 1 to ITERATIONS with a step of 250 iterations.
    
    Note: Setting up the random seed, the execution can take a while (around 1 hour for each experiment), so be patient. For impatients, we also present additional results for a small-quick comparison using the -e option for the excerpt version of the model. To understand this, read the final note at the end of this document.
    
    Change the -m parameter to flat for the flat Monte Carlo method, Greedy for the Greedy MCTS, and random for Random Sampling. For the Random Sampling, in the case of using the complete version of the feature model (10e9 configurations) or other large-scale feature models, you need to use the BDDSampler by Heradio et al. The integration of BDDSampler within our framework is out of the scope of this work. Thus, there is no script at this moment to automate the random sampling results from the BDDSampler. In the case of using the excerpt version (-e parameter), the maximum number of iterations is equal to the maximum number of configurations of the feature model.
  - For the jHipster feature model):
    
    python main_comparison_jhipster.py -it 1000 -s 2021 -m MCTS
    
    Use the same parameters and comments as the previous script hold. However, the Random Sampling strategy can be directly used for the jHipster feature model because all configurations are available.
  Note about randomness in Monte Carlo methods
  
  The underlying principle of operation of Monte Carlo methods is to use randomness to solve problems. Our framework relies on the Python random module, concretely we use the choice, shuffle, and sample methods to implement the selection and simulation steps of the MCTS method, as well as the possible successors of the states and possible actions of the problems. To provide reproducibility in our framework, we use a random seed initialized at the beginning of the experiment.
  
  However, the random module is not the only source of randomness in our framework. The MCTS method highly works with data structures (e.g., the search tree) that do not maintain the order of the states (e.g., sets, maps, or dictionaries). For instance, in a configuration of a feature model, the order of the features is irrelevant. Using those structures does not guarantee obtaining identical results when using methods like random.choice. Moreover, states in our framework can represent features, configurations of the feature model, or even feature models like in the reverse engineering problem. Maintaining a total order for those concepts is not straightforward. For example, defining when a feature model is lesser than others is not trivial. This change also impacts, and significatively degrades the performance of the solution because it requires continuously sorting the collections or using inefficient sorted data structures, which Monte Carlo methods do not really need. That is, there is an important trade-off between performance and reproducibility when dealing with Monte Carlo methods and randomness that should be considered.
  
  To alleviate these issues and provide maximum reproducibility, we have modified our framework to use sorted data structures in all cases, defining when necessary a total order between the states. Despite this, some experiments can still present a slight variation. This is due to draws in the sorted elements (e.g., features with the same names). However, these variations do not affect the overall results and conclusions of our research.
  
  References

Name		Name	Last commit message	Last commit date
Latest commit History 169 Commits
additional experiments		additional experiments
img		img
models		models
montecarlo_framework		montecarlo_framework
results		results
tests		tests
.gitignore		.gitignore
README.md		README.md
experiments_scripts.sh		experiments_scripts.sh
main_complete_partial_config.py		main_complete_partial_config.py
main_find_defective_config.py		main_find_defective_config.py
main_optimize_config.py		main_optimize_config.py
main_reverse_engineering.py		main_reverse_engineering.py
pytest.ini		pytest.ini
requirements.txt		requirements.txt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

A Monte Carlo Tree Search conceptual framework for feature model analyses

Table of Contents

Artifact description

The Monte Carlo conceptual framework

Interfaces

Algorithms and Monte Carlo methods

Configuration

Knowledge inference

Installation and execution

Requirements

Download and setup

Analyzing problems with the Monte Carlo framework

Results

Experiment replication

Replicating results from problem analyses

Replicating results from the comparison of Monte Carlo methods

Note about randomness in Monte Carlo methods

References

About

Releases 2

Packages

Contributors 5

Languages

diverso-lab/fm_montecarlo

Folders and files

Latest commit

History

Repository files navigation

A Monte Carlo Tree Search conceptual framework for feature model analyses

Table of Contents

Artifact description

The Monte Carlo conceptual framework

Interfaces

Algorithms and Monte Carlo methods

Configuration

Knowledge inference

Installation and execution

Requirements

Download and setup

Analyzing problems with the Monte Carlo framework

Results

Experiment replication

Replicating results from problem analyses

Replicating results from the comparison of Monte Carlo methods

Note about randomness in Monte Carlo methods

References

About

Topics

Resources

Stars

Watchers

Forks

Releases 2

Packages 0

Contributors 5

Languages

Packages