SustainDC (DCRL-Green) - Benchmarking for Sustainable Data Center Control

Table of Contents

1. Introduction
2. Features
3. Installation
4. Quick Start Guide
5. Environment Details
6. Customization
7. Benchmarking Algorithms
8. Evaluation Metrics
9. Dashboard
10. Contributing
11. Contact
12. Star History
13. License
14. Citation

Disclaimer

This work builds on our previous research and extends the methodologies and insights gained from our previous work. The original code, referred to as DCRL-Green, can be found in the legacy branch of this repository. For users looking for the original implementation, please visit the legacy branch. The current repository, SustainDC, represents an advanced iteration of DCRL-Green, incorporating enhanced features and improved benchmarking capabilities. This evolution reflects our ongoing commitment to advancing sustainable data center control. Consequently, the repository name remains dc-rl to maintain continuity with our previous work.

Introduction

SustainDC is a set of Python environments for benchmarking multi-agent reinforcement learning (MARL) algorithms in data centers (DC). It focuses on sustainable DC operations, including workload scheduling, cooling optimization, and auxiliary battery management. This repository contains the code and datasets for the paper SustainDC: Benchmarking for Sustainable Data Center Control.

This codebase is for the paper: SustainDC: Benchmarking for Sustainable Data Center Control, which is under review for the NeurIPS 2024 Datasets and Benchmarks Track.

Documentation and Installation

Refer to the docs for broader documentation of SustainDC.

Features

• Highly Customizable Environments: Allows users to define and modify various aspects of DC operations, including server configurations, cooling systems, and workload traces.
• Multi-Agent Support: Enables the testing of MARL controllers with both homogeneous and heterogeneous agents, facilitating the study of collaborative and competitive strategies in DC management.
• Gymnasium Integration: Environments are wrapped in the Gymnasium Env class, making it easy to benchmark different control strategies using standard reinforcement learning libraries.
• Realistic External Variables: Incorporates real-world data such as weather conditions, carbon intensity, and workload traces to simulate the dynamic and complex nature of DC operations.
• Collaborative Reward Mechanisms: Supports the design of custom reward structures to promote collaborative optimization across different DC components.
• Benchmarking Suite: Includes scripts and tools for evaluating the performance of various MARL algorithms, providing insights into their effectiveness in reducing energy consumption and carbon emissions.

Installation

Prerequisites

• Python 3.10+
• Conda (for creating virtual environments)
• Dependencies listed in requirements.txt

Steps

1. Clone the repository:

   git clone https://github.com/HewlettPackard/dc-rl.git sustaindc

2. Navigate to the repository directory:

   cd sustaindc

3. Create a new Conda environment with Python 3.10:

   conda create --name sustaindc python=3.10

4. Activate the newly created environment:

   conda activate sustaindc

5. Install the required packages:

   pip install -r requirements.txt

Quick Start Guide

1. Setup Configuration: Customize the dc_config.json file to specify your DC environment settings.

2. Environment Configuration: The main environment for wrapping the environments is sustaindc_env.py, which reads configurations from env_config and manages the external variables using managers for weather, carbon intensity, and workload.

3. Train Example: Specify location inside harl.config.envs_cfgs.sustaindc.yaml. Specify other algorithm hyperparameteres in harl.config.algos_cfgs.<algorith>.yaml.

   python train_sustaindc.py --algo happo --exp_name happo 

4. Monitor training on Tensorboard

   tensorboard --logdir /results/sustaindc/<location>/happo

5. Evaluation Example:

   python eval_sustaindc.py

   The results are stored in the SAVE_EVAL(can be configured inside eval_harl.py with other parameters like choice of checkpoint, location, run etc) folder.

Environment Details

SustainDC consists of three interconnected environments that simulate various aspects of data center operations: the Workload Environment, the Data Center Environment, and the Battery Environment. These environments work together to provide a comprehensive platform for benchmarking MARL algorithms aimed at optimizing energy consumption and reducing the carbon footprint of DCs.

Workload Environment

The Workload Environment manages the execution and scheduling of delayable workloads within the data center. It simulates the computational demand placed on the data center by using real-world workload traces from sources like Alibaba and Google.

Key Features

• Observation Space:

  • Time of Day and Year: Provides a periodic understanding of time using sine and cosine representations.
  • Grid Carbon Intensity (CI): Includes current and forecasted carbon intensity values to help the agent optimize workload scheduling based on carbon emissions.
  • Rescheduled Workload Left: Tracks the amount of workload that has been rescheduled but not yet executed.

• Action Space:

  • Store Delayable Tasks: Allows the agent to store delayable tasks for future execution.
  • Compute All Immediate Tasks: Enables the agent to process all current tasks immediately.
  • Maximize Throughput: Balances immediate and delayed tasks based on the current carbon intensity.

Data Center Environment

The Data Center Environment models the IT and HVAC systems of a data center, focusing on optimizing energy consumption and cooling. It simulates the electrical and thermal behavior of the DC components, including servers and cooling systems.

Key Features

• Observation Space:

  • Time of Day and Year: Provides a periodic understanding of time using sine and cosine representations.
  • Ambient Weather (Dry Bulb Temperature): Current outside temperature affecting the cooling load.
  • IT Room Temperature: Current temperature inside the data center, crucial for maintaining optimal server performance.
  • Previous Step Energy Consumption: Historical data on cooling and IT energy consumption for trend analysis.
  • Grid Carbon Intensity (CI): Forecasted carbon intensity values to optimize cooling strategies.

• Action Space:

  • Decrease Setpoint: Lowers the cooling setpoint to increase cooling, consuming more energy for cooling but reducing IT energy consumption.
  • Maintain Setpoint: Keeps the current cooling setpoint constant.
  • Increase Setpoint: Raises the cooling setpoint to reduce cooling energy consumption but increases IT energy consumption.

Battery Environment

The Battery Environment simulates the charging and discharging cycles of batteries used in the DC. It models how batteries can be charged from the grid during periods of low carbon intensity and provide auxiliary energy during periods of high carbon intensity.

Key Features

• Observation Space:

  • Time of Day and Year: Provides a periodic understanding of time using sine and cosine representations.
  • State of Charge (SoC): Current energy level of the battery.
  • Grid Energy Consumption: Combined energy consumption of IT and cooling systems.
  • Grid Carbon Intensity (CI): Current and forecasted carbon intensity values to determine optimal charging and discharging times.

• Action Space:

  • Charge Battery: Stores energy in the battery during periods of low carbon intensity.
  • Hold Energy: Maintains the current state of charge.
  • Discharge Battery: Provides auxiliary energy to the data center during periods of high carbon intensity.

Interconnected Environments

These three environments are interconnected to simulate realistic DC operations:

• The Workload Environment generates the computational demand that the Data Center Environment must process. This includes managing the scheduling of delayable tasks to optimize energy consumption and reduce the carbon footprint.

• The Data Center Environment handles the cooling and IT operations required to process the workloads. Higher computational demand results in increased heat generation, necessitating more cooling and energy consumption.

• The Battery Environment supports the DC by providing auxiliary energy during periods of high carbon intensity, helping to reduce the overall carbon footprint. It is affected by both the Workload Environment and the Data Center Environment. The workload impacts heat generation, which in turn affects the cooling requirements and energy consumption, influencing the battery's charging and discharging cycles.

Together, these interconnected environments provide a dynamic platform for benchmarking MARL algorithms, helping to develop strategies for more sustainable and efficient DC operations.

External Input Data

SustainDC uses external input data to provide a realistic simulation environment:

Workload

The Workload external data in SustainDC represents the computational demand placed on the DC. By default, SustainDC includes a collection of open-source workload traces from Alibaba and Google DCs. Users can customize this component by adding new workload traces to the data/Workload folder or specifying a path to existing traces in the sustaindc_env.py file under the workload_file configuration.

Weather

The Weather external data in SustainDC captures the ambient environmental conditions impacting the DC's cooling requirements. By default, SustainDC includes weather data files in the .epw format from various locations where DCs are commonly situated. These locations include Arizona, California, Georgia, Illinois, New York, Texas, Virginia, and Washington. Users can customize this component by adding new weather files to the data/Weather folder or specifying a path to existing weather files in the sustaindc_env.py file under the weather_file configuration.

Each .epw file contains hourly data for various weather parameters, but for our purposes, we focus on the ambient temperature.

Carbon Intensity

The Carbon Intensity (CI) external data in SustainDC represents the carbon emissions associated with electricity consumption. By default, SustainDC includes CI data files for various locations: Arizona, California, Georgia, Illinois, New York, Texas, Virginia, and Washington. These files are located in the data/CarbonIntensity folder and are extracted from https://api.eia.gov/bulk/EBA.zip. Users can customize this component by adding new CI files to the data/CarbonIntensity folder or specifying a path to existing files in the sustaindc_env.py file under the cintensity_file configuration.

Furthermore, in the figure below, we show the average daily carbon intensity against the average daily coefficient of variation (CV) for various locations. This figure highlights an important perspective on the variability and magnitude of carbon intensity values across different regions. Locations with a high CV indicate greater fluctuation in carbon intensity, offering more "room to play" for DRL agents to effectively reduce carbon emissions through dynamic actions. Additionally, locations with a high average carbon intensity value present greater opportunities for achieving significant carbon emission reductions. The selected locations are highlighted, while other U.S. locations are also plotted for comparison. Regions with both high CV and high average carbon intensity are identified as prime targets for DRL agents to maximize their impact on reducing carbon emissions.

Below is a summary of the selected locations, typical weather values, and carbon emissions characteristics:

Location	Typical Weather	Carbon Emissions
Arizona	Hot, dry summers; mild winters	High avg CI, High variation
California	Mild, Mediterranean climate	Medium avg CI, Medium variation
Georgia	Hot, humid summers; mild winters	High avg CI, Medium variation
Illinois	Cold winters; hot, humid summers	High avg CI, Medium variation
New York	Cold winters; hot, humid summers	Medium avg CI, Medium variation
Texas	Hot summers; mild winters	Medium avg CI, High variation
Virginia	Mild climate, seasonal variations	Medium avg CI, Medium variation
Washington	Mild, temperate climate; wet winters	Low avg CI, Low variation

Customization

SustainDC offers extensive customization options to tailor the environments to specific needs and configurations. Users can modify various parameters and components across the Workload, Data Center, and Battery environments, as well as external variables like weather carbon intensity data, and workload trace.

Main Configuration File

The main environment for wrapping the environments is sustaindc_env.py, which reads configurations from dc_config.json and manages the external variables using managers for weather, carbon intensity, and workload.

Example Configuration

env_config = {
    # Agents active
    'agents': ['agent_ls', 'agent_dc', 'agent_bat'],

    # Datafiles
    'location': 'ny',
    'cintensity_file': 'NYIS_NG_&_avgCI.csv',
    'weather_file': 'USA_NY_New.York-Kennedy.epw',
    'workload_file': 'Alibaba_CPU_Data_Hourly_1.csv',

    # Data Center maximum capacity
    'datacenter_capacity_mw': 1,
    
    # Battery capacity
    'max_bat_cap_Mw': 2,
    
    # Collaborative weight in the reward
    'individual_reward_weight': 0.8,
    
    # Flexible load ratio
    'flexible_load': 0.4,
    
    # Specify reward methods
    'ls_reward': 'default_ls_reward',
    'dc_reward': 'default_dc_reward',
    'bat_reward': 'default_bat_reward'
}

Data Center Configuration File

The customization of the DC is done through the dc_config.json file located in the utils folder. This file allows users to specify every aspect of the DC environment design.

Example Configuration File Structure

{
    "data_center_configuration": {
        "NUM_ROWS": 4,
        "NUM_RACKS_PER_ROW": 5,
    },
    "hvac_configuration": {
        "C_AIR": 1006,
        "RHO_AIR": 1.225,
        "CRAC_SUPPLY_AIR_FLOW_RATE_pu": 0.00005663,
        "CRAC_REFRENCE_AIR_FLOW_RATE_pu": 0.00009438,
        "CRAC_FAN_REF_P": 150,
        "CHILLER_COP_BASE": 5.0,
        "CHILLER_COP_K": 0.1,
        "CHILLER_COP_T_NOMINAL": 25.0,
        "CT_FAN_REF_P": 1000,
        "CT_REFRENCE_AIR_FLOW_RATE": 2.8315,
        "CW_PRESSURE_DROP": 300000,
        "CW_WATER_FLOW_RATE": 0.0011,
        "CW_PUMP_EFFICIENCY": 0.87,
        "CT_PRESSURE_DROP": 300000,
        "CT_WATER_FLOW_RATE": 0.0011,
        "CT_PUMP_EFFICIENCY": 0.87
    },
    "server_characteristics": {
        "CPU_POWER_RATIO_LB": [0.01, 1.00],
        "CPU_POWER_RATIO_UB": [0.03, 1.02],
        "IT_FAN_AIRFLOW_RATIO_LB": [0.01, 0.225],
        "IT_FAN_AIRFLOW_RATIO_UB": [0.225, 1.0],
    }
}

Adding New Workload Data

Overview

By default, SustainDC includes workload traces from Alibaba and Google DC. These traces are used to simulate the tasks that the DC needs to process, providing a realistic and dynamic workload for benchmarking purposes.

Data Source

The default workload traces are extracted from:

• Alibaba 2017 CPU Data (https://github.com/alibaba/clusterdata)
• Google 2011 CPU Data (https://github.com/google/cluster-data)

Expected File Format

Workload trace files should be in CSV format, with two columns: a timestamp or index (must be unnamed), and the corresponding DC Utilization (cpu_load). The CPU load must be expressed as a fraction of the DC utilization (between 0 and 1). The workload file must contain one year of data with an hourly periodicity (365*24=8760 rows).

Example Workload Trace File

,cpu_load
1,0.380
2,0.434
3,0.402
4,0.485
...

Integration Steps

1. Place the new workload trace file in the data/Workload folder.
2. Update the workload_file entry in env_config with the path to the new workload trace file.

Adding New Carbon Intensity Data

Overview

Carbon Intensity (CI) data represents the carbon emissions associated with electricity consumption. SustainDC includes CI data files for various locations to simulate the carbon footprint of the DC's energy usage.

Data Source

The default carbon intensity data files are extracted from:

• U.S. Energy Information Administration (EIA) API: https://api.eia.gov/bulk/EBA.zip

Expected File Format

Carbon intensity files should be in CSV format, with columns representing different energy sources and the average carbon intensity. The columns typically include:

• timestamp: The time of the data entry.
• WND: Wind energy.
• SUN: Solar energy.
• WAT: Water (hydropower) energy.
• OIL: Oil energy.
• NG: Natural gas energy.
• COL: Coal energy.
• NUC: Nuclear energy.
• OTH: Other energy sources.
• avg_CI: The average carbon intensity.

Example Carbon Intensity File

timestamp,WND,SUN,WAT,OIL,NG,COL,NUC,OTH,avg_CI
2022-01-01 00:00:00+00:00,1251,0,3209,0,15117,2365,4992,337,367.450
2022-01-01 01:00:00+00:00,1270,0,3022,0,15035,2013,4993,311,363.434
2022-01-01 02:00:00+00:00,1315,0,2636,0,14304,2129,4990,312,367.225
2022-01-01 03:00:00+00:00,1349,0,2325,0,13840,2334,4986,320,373.228
...

Integration Steps

1. Add the new carbon intensity data files to the data/CarbonIntensity folder.
2. Update the cintensity_file entry in env_config with the path to the new carbon intensity file.

Adding New Weather Data

Overview

Weather data captures the ambient environmental conditions that impact the DC's cooling requirements. SustainDC includes weather data files in the .epw format from various locations where DCs are commonly situated.

Data Source

The default weather data files are extracted from:

• EnergyPlus Weather Data: https://energyplus.net/weather

Expected File Format

Weather data files should be in the .epw (EnergyPlus Weather) format, which includes hourly data for various weather parameters such as temperature, humidity, wind speed, etc.

Example Weather Data File

Below is a partial example of an .epw file:

LOCATION,New York JFK Int'l Ap NY USA,USA,NY,716070,40.63,-73.78,-5.0,3.4
DESIGN CONDITIONS,1,New York JFK Int'l Ap Ann Clg .4% Condns DB=>MWB,25.0
...
DATA PERIODS,1,1,Data,Sunday, 1/ 1,12/31
...
2022,1,1,1,60.1,45.1,1004.1,0,0,4.1,80.1,30.0
2022,1,1,2,59.8,44.9,1004.0,0,0,4.1,80.0,29.8
...

Integration Steps

1. Add the new weather data files in the .epw format to the data/Weather folder.
2. Update the weather_file entry in env_config with the path to the new weather file.

Custom Reward Structures

SustainDC allows users to define custom reward structures to promote collaborative optimization across different DC components. Users can modify the reward functions in the utils/reward_creator.py file to suit their specific optimization goals.

By leveraging these customization options, users can create highly specific and optimized simulations that reflect the unique requirements and challenges of their DC operations.

Benchmarking Algorithms

SustainDC supports a variety of reinforcement learning algorithms for benchmarking. This section provides an overview of the supported algorithms and highlights their differences.

Supported Algorithms

Algorithm	Description	Use Case
PPO (Proximal Policy Optimization)	Balances simplicity and performance with a clipped objective function for stable training.	Single-agent environments.
IPPO (Independent PPO)	Each agent operates independently with its own policy and value function.	Multi-agent systems with distinct roles.
MAPPO (Multi-Agent PPO)	Uses a centralized value function for better coordination among agents.	Cooperative multi-agent tasks.
HAPPO (Heterogeneous Agent PPO)	Designed for heterogeneous agents with different observation and action spaces.	Complex environments with diverse agents.
HATRPO (Heterogeneous Agent Trust Region Policy Optimization)	Adapts TRPO for heterogeneous multi-agent settings with stability and robustness.	Complex environments requiring robust policy updates.
HAA2C (Heterogeneous Agent Advantage Actor-Critic)	Extends A2C to multi-agent settings with individual actor and critic networks.	Scenarios with different types of observations and actions.
HAD3QN (Heterogeneous Agent Dueling Double Deep Q-Network)	Combines dueling networks and double Q-learning for stability and performance.	Environments needing fine-grained action distinctions.
HASAC (Heterogeneous Agent Soft Actor-Critic)	Uses entropy regularization for exploration in multi-agent settings.	Adapted to discrete action spaces and high exploration needs.

Differences and Use Cases

• PPO vs. IPPO: PPO is for single-agent setups, while IPPO suits multi-agent environments with independent learning.
• IPPO vs. MAPPO: IPPO treats agents independently; MAPPO coordinates agents with a centralized value function, ideal for cooperative tasks.
• MAPPO vs. HAPPO: Both use centralized value functions, but HAPPO is for heterogeneous agents with different capabilities.
• HAPPO vs. HATRPO: HAPPO uses PPO-based updates; HATRPO adapts TRPO for more stable and robust policy updates in heterogeneous settings.
• HAPPO vs. HAA2C: HAPPO is PPO-based; HAA2C extends A2C to multi-agent systems, offering stability and performance trade-offs.
• HAA2C vs. HAD3QN: HAA2C is an actor-critic method; HAD3QN uses value-based learning with dueling and double Q-learning
• HAD3QN vs. HASAC: HAD3QN is value-based; HASAC uses entropy regularization for environments with continuous action spaces.

By supporting a diverse set of algorithms, SustainDC allows researchers to benchmark and compare the performance of various reinforcement learning approaches in the context of sustainable DC control.

Running Benchmarks

Benchmarking the performance of different reinforcement learning algorithms using SustainDC involves several steps. This section provides a detailed explanation of the benchmarking process.

1. Configuring the Environment:

   Before running the benchmarks, ensure that the environment is properly configured. Customize the dc_config.json file to specify your DC environment settings, and update the environment configurations in harl.config.envs_cfgs.sustaindc.yaml.

2. Setting Algorithm Parameters:

   Each algorithm has specific hyperparameters that need to be set before running the training script. Modify the parameters in the corresponding YAML configuration file located in the harl/configs/algos_cfgs directory. For example, to set the parameters for the HAPPO algorithm, edit the harl/configs/algos_cfgs/happo.yaml file.

3. Run the Training Script:

   Once the environment and algorithm parameters are configured, you can run the training script. Use the following command, replacing <algorithm> with the desired algorithm and <experiment_name> with a name for the experiment:

   python train_sustaindc.py --algo <algorithm> --exp_name <experiment_name>

   For example, to run the training using the HAPPO algorithm and name the experiment happo, you would use:

   python train_sustaindc.py --algo happo --exp_name happo_experiment

4. Monitor Training with TensorBoard:

   To visualize the progress of your experiments, you can use TensorBoard. Run the following command in another cell or terminal:

   %tensorboard --logdir results/sustaindc --port 6006

   This command will start a TensorBoard server on PORT 6006, and you can monitor the training metrics and visualizations in real-time. You can use other unused PORT. Open your web browser and navigate to http://localhost:<PORT> to view real-time metrics and visualizations of the training process.

5. Evaluate the Trained Results:

   After training is complete, evaluate the performance of the trained models using the evaluation script. Make sure to specify the correct RUN to the trained experiment in the eval_sustaindc.py file:

   python eval_sustaindc.py

Evaluation Metrics

• Carbon Footprint (CFP): Cumulative carbon emissions over the evaluation period.
• HVAC Energy: Energy consumed by the DC cooling system.
• IT Energy: Energy consumed by the DC servers.
• Water Usage: Efficient utilization of water for cooling.
• Task Dropped: Number of dropped tasks due to workload scheduling.

Dashboard

To get an in-depth look at the SustainDC dashboard and see real-time metrics, watch the video demonstration. The video showcases the dynamic plotting of variables from the agents, environments, and metrics, providing a comprehensive view of the DC operations.

Click on the screenshot below to watch the video (right-click and select "Open link in new tab" to view in a new tab):

In the video, you will see:

• Real-time plotting of agent variables: Watch how the agents' actions and states are visualized dynamically.
• Environment metrics: Observe the DC's performance metrics, including energy consumption, cooling requirements, and workload distribution.
• Interactive dashboard features: Learn about the various interactive elements of the dashboard that allow for detailed analysis and monitoring.

If you wish to download the video directly, click here.

Contributing

We welcome contributions from the community! Whether it's bug fixes, new features, or improvements to the documentation, your help is appreciated. Please follow the guidelines below to contribute to SustainDC.

How to Contribute

1. Fork the Repository: Fork the SustainDC repository to your own GitHub account.

2. Clone the Repository: Clone the forked repository to your local machine:

   git clone https://github.com/your-username/dc-rl.git

3. Create a Branch: Create a new branch for your feature or bug fix:

    git checkout -b feature-or-bugfix-name

4. Make Changes: Make your changes to the codebase.

5. Commit Changes: Commit your changes with a clear and descriptive commit message:

   git commit -m "Description of your changes"

6. Push Changes: Push your changes to your forked repository:

   git push origin feature-or-bugfix-name

7. Create a Pull Request: Open a pull request on the original SustainDC repository. Provide a clear description of what your changes do and any relevant information for the review process.

Code of Conduct

Please note that we have a Code of Conduct. By participating in this project, you agree to abide by its terms.

Reporting Issues

If you find a bug or have a feature request, please create an issue on the GitHub Issues page. Provide as much detail as possible to help us understand and address the issue.

Guidelines

• Ensure your code follows the project's coding standards and conventions.
• Write clear, concise commit messages.
• Update documentation if necessary.
• Add tests to cover your changes if applicable.

Thank you for contributing to SustainDC!

Contact

If you have any questions, suggestions, or issues, please feel free to contact us. We are here to help and support you in using SustainDC.

Contact Information

• Email: soumyendu.sarkar@hpe.com
• GitHub Issues: GitHub Issues Page

Star History

License

MIT License

The majority of this project is licensed under the MIT License. See the LICENSE file for more details.

Creative Commons Attribution-NonCommercial 4.0 International License

Some parts of this project are derived from code originally published under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) License. The original authors retain the copyright to these portions, and they are used here under the terms of the CC BY-NC 4.0 license.

Attribution for CC BY-NC 4.0 Licensed Material

• Original Repository

Combined Work

This combined work is available under both the MIT License for the new contributions and the CC BY-NC 4.0 License for the portions derived from the original work.

Citation

If you find our paper or this repository helpful in your research or project, please consider citing our work using the following BibTeX citation:

@article{Sarkar2024Mar,
	author = {Sarkar, Soumyendu and Naug, Avisek and Luna, Ricardo and Guillen, Antonio and Gundecha, Vineet and Ghorbanpour, Sahand and Mousavi, Sajad and Markovikj, Dejan and Babu, Ashwin Ramesh},
	title = {{Carbon Footprint Reduction for Sustainable Data Centers in Real-Time}},
	journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
	volume = {38},
	number = {20},
	pages = {22322--22330},
	year = {2024},
	month = mar,
	issn = {2374-3468},
	doi = {10.1609/aaai.v38i20.30238}
}