Self-Hosted Agent-Based Modeling Platforms: NetLogo vs GAMA vs Repast Simphony

Introduction

Agent-based modeling (ABM) is a computational approach that simulates the actions and interactions of autonomous agents — individual entities with defined behaviors — to understand emergent system-level patterns. From epidemiological modeling and traffic simulation to ecosystem dynamics and social behavior analysis, ABM platforms enable researchers to explore complex systems that resist traditional equation-based approaches.

This guide compares three leading open-source agent-based modeling platforms that can be deployed on self-hosted HPC infrastructure: NetLogo (Northwestern University), GAMA Platform (IRD/INRAE), and Repast Simphony (Argonne National Laboratory). We examine installation, scalability, modeling capabilities, and suitability for different research domains.

Why Self-Host Your Agent-Based Modeling?

Running ABM simulations on your own infrastructure provides distinct advantages for research integrity and computational efficiency. Agent-based models are inherently parameter-intensive — a single simulation may need to run thousands of times with varying parameters to produce statistically meaningful results. Cloud-based simulation services charge per compute-hour, making large parameter sweeps prohibitively expensive. Self-hosted HPC clusters with job schedulers like SLURM and PBS can execute hundreds of parallel simulation runs simultaneously at no incremental cost.

Second, model transparency and reproducibility are fundamental to scientific computing. When you self-host your ABM platform, you control the exact software version, random seed, and computational environment. This eliminates the “works on my machine” problem that plagues simulation-based research. For complex models involving peer review and replication, this environmental control is essential.

Third, many ABM applications involve sensitive or proprietary data — epidemiological simulations with patient data, defense modeling, urban planning with census data, or financial market simulations. Self-hosting keeps this data within your institutional firewall, addressing security and compliance requirements that cloud platforms cannot satisfy. For related modeling approaches, see our scientific simulation comparison.

Tool Comparison Overview

NetLogo

NetLogo is the most widely adopted ABM platform, used extensively in education and research with over 1,100 GitHub stars. Its domain-specific language is designed for accessibility — researchers without programming backgrounds can build and run models within hours of starting.

Docker Installation

1
2
3
4
5
6
7

# Pull the official COMSES Docker image
docker pull comses/netlogo:6.4.0
docker run --rm -v /data:/data comses/netlogo:6.4.0 \
  --headless \
  --model /data/models/Flocking.nlogo \
  --experiment experiment \
  --table /data/output/table-output.csv

Headless Batch Execution

NetLogo’s BehaviorSpace module enables parameter sweeping through headless execution:

1
2
3
4
5
6
7

# Run parameter sweep
java -Xmx8192M -cp NetLogo.jar \
  org.nlogo.headless.Main \
  --model MyModel.nlogo \
  --experiment "sweep-params" \
  --table results.csv \
  --threads 16

Key Strengths

NetLogo’s greatest strength is its ecosystem and community. The NetLogo Models Library contains over 200 pre-built models spanning biology, physics, chemistry, social science, and computer science. These models serve as both teaching examples and starting points for custom research. The extensive documentation, active mailing list, and published literature make NetLogo the easiest platform to start with.

The NetLogo extensions API enables integration with external libraries: the R extension allows statistical analysis within simulations, the Python extension enables machine learning integration, and the GIS extension provides basic geospatial support. For research groups new to ABM, NetLogo’s gentle learning curve and rich ecosystem make it the best entry point.

NetLogo 3D adds volumetric spatial modeling — essential for particle simulations, building evacuation modeling, and biological tissue simulations. For neuroscience simulation applications, see our neuroscience simulation guide.

GAMA Platform

GAMA (GIS and Agent-based Modeling Architecture) is purpose-built for spatially explicit agent-based modeling. Unlike NetLogo’s optional GIS extension, GAMA’s entire modeling framework is built around geospatial data from the ground up.

Docker Installation

 1
 2
 3
 4
 5
 6
 7
 8
 9
10

# Build from source for headless HPC deployment
git clone https://github.com/gama-platform/gama.git
cd gama
docker build -t gama-platform:latest .

# Run headless GAMA experiment
docker run --rm -v /data:/data gama-platform:latest \
  /usr/local/gama/headless/gama-headless.sh \
  /data/models/traffic.gaml \
  /data/output/

GAML Modeling Language

GAMA uses GAML (GAma Modeling Language), a rich domain-specific language that treats GIS data as first-class citizens:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

model TrafficSimulation

global {
    file shapefile_buildings <- file("../data/buildings.shp");
    file shapefile_roads <- file("../data/roads.shp");
    geometry shape <- envelope(shapefile_buildings);
}

species vehicle skills: [moving] {
    float speed <- 50.0 + rnd(20.0);
    reflex move {
        do goto target: roads;
    }
}

experiment Traffic type: gui {
    output {
        display map type: opengl {
            species vehicle aspect: base;
            graphics "buildings" {
                draw shapefile_buildings color: #gray;
            }
        }
    }
}

Key Strengths

GAMA’s native GIS integration is unmatched among ABM platforms. It supports shapefiles, GeoJSON, OSM data, raster grids, and network files for graph-based road and utility networks. For urban planning, evacuation modeling, epidemiology, and transportation research, GAMA’s spatial capabilities dramatically reduce development time.

The platform’s multi-level modeling supports agents nested within agents nested within environments — cities contain neighborhoods, neighborhoods contain buildings, buildings contain people — each level with independent behaviors. This hierarchical approach maps naturally to urban systems, ecological food webs, and organizational modeling.

Repast Simphony

Repast Simphony, developed at Argonne National Laboratory, is the most computationally sophisticated of the three platforms. Written in Java, it targets research groups that need maximum performance, distributed computing, and integration with existing Java scientific libraries.

Installation

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

# Clone and build
git clone https://github.com/Repast/repast.simphony.git
cd repast.simphony
./gradlew build

# Run headless batch
java -cp repast.simphony.bin.jar \
  -Xmx16G \
  repast.simphony.batch.BatchMain \
  -params params.xml \
  /data/models/MyModel.rs

ReLogo DSL

Repast offers ReLogo — a Logo-inspired domain-specific language that provides NetLogo-like syntax on top of Repast’s high-performance Java engine:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19

;; ReLogo example - predator-prey model
globals [grass-regrowth-time]

breed [wolves wolf]
breed [sheep a-sheep]

to setup
  clear-all
  set grass-regrowth-time 20
  create-wolves 50
  create-sheep 200
  ask sheep [setxy random-xcor random-ycor]
end

to go
  ask wolves [hunt]
  ask sheep [graze reproduce]
  tick
end

Key Strengths

Repast’s defining advantage is distributed execution. The Repast for High Performance Computing (Repast HPC) variant runs agent-based models across MPI-enabled clusters, enabling simulations with millions of agents that exceed single-machine memory limits. For epidemiological models tracking individual-level disease spread across populations, or social simulation with millions of interacting agents, Repast HPC is the only viable open-source option.

Repast also provides the most mature Java API, allowing seamless integration with libraries like Apache Commons Math, Weka, and JFreeChart. For research groups already invested in the Java scientific ecosystem, Repast provides the smoothest development experience.

For robotics and multi-agent systems with physical embodiment, see our robotics simulation guide.

Choosing the Right ABM Platform

• Choose NetLogo for education, rapid prototyping, and research where accessibility matters more than raw performance. The extensive model library and community support make it the best starting point for ABM newcomers.

• Choose GAMA when your model is inherently spatial — urban systems, transportation networks, environmental modeling, or anything involving GIS data. GAMA’s native geospatial support eliminates the friction of bolting GIS onto a non-spatial platform.

• Choose Repast for large-scale, computationally intensive simulations requiring distributed execution across HPC clusters. If your model has millions of agents, uses complex Java libraries, or needs MPI-based parallelization, Repast is the tool of choice.

FAQ

Which ABM platform is best for beginners?

NetLogo is the clear choice for beginners. Its Logo-based language is designed specifically for accessibility, the built-in Models Library provides over 200 working examples, and the visual interface allows you to see agent behavior in real-time. Most researchers can build their first working model within a day of starting with NetLogo.

Can I run these platforms without a graphical interface?

Yes, all three support headless execution. NetLogo has a dedicated headless mode for BehaviorSpace experiments. GAMA provides gama-headless.sh for batch execution on HPC clusters. Repast supports headless batch mode through its BatchMain class and distributed execution via Repast HPC.

How do these platforms handle large numbers of agents?

NetLogo handles up to approximately 100,000 agents comfortably on a single machine before performance degrades. GAMA uses optimized spatial indexing and can scale to roughly 500,000 agents on adequate hardware. Repast HPC can simulate millions of agents across distributed MPI clusters — an order of magnitude beyond single-machine platforms.

What statistical analysis tools integrate with these platforms?

NetLogo integrates with R through its R extension and Python through its Python extension for statistical analysis within simulations. GAMA integrates with R through its R plugin. Repast integrates with any Java-compatible statistics library including Apache Commons Math, JFreeChart, and Weka natively through its Java API.

Are ABM simulations deterministic?

All three platforms use pseudo-random number generators with configurable seeds, so simulations are reproducible when the same seed is used. NetLogo and GAMA both provide random seed configuration in their experiment setup. Repast uses the Mersenne Twister PRNG with configurable seeding. For stochastic models, multiple runs with different seeds are required for statistical validity.

Can I export simulation results for analysis?

Yes, all platforms support data export. NetLogo BehaviorSpace produces CSV tables of parameter sweeps. GAMA supports CSV, shapefile, and JSON export. Repast offers programmatic data export through its Java API plus built-in charting capabilities.

💰 想测试你的市场判断力？我用 Polymarket 做预测市场交易——这是全球最大的预测市场平台，从大选结果到技术监管时间线，什么都可以押注。和赌博不同，这是真正的信息市场：你懂的信息越多，胜率越高。我靠预测技术相关事件的走向已经赚了不少。用我的邀请链接注册：Polymarket.com