AI and Artificial-Life Tools on Linux

Linux has always been the natural habitat for AI and computational biology research. Whether you are running modern deep-learning workloads on NVIDIA hardware, evolving neural architectures with genetic algorithms, or simulating artificial ecosystems, every serious open-source framework runs best—or only—on Linux. This guide walks through setting up and validating the major stacks, from classic a-life simulators to PyTorch and JAX, with concrete commands for each major distro family.

Prerequisites and System Baseline

Before installing anything, confirm your kernel and driver situation. GPU-accelerated frameworks require a relatively recent kernel (5.15+ is fine; 6.x is better) and, for NVIDIA cards, the proprietary driver. AMD GPUs use the open ROCm stack. CPU-only work needs nothing special beyond a 64-bit install and Python 3.10+.

uname -r
python3 --version
lscpu | grep -E 'Model name|CPU\(s\)|Thread'
nvidia-smi   # NVIDIA only — will fail cleanly if no GPU present

Python Environment Strategy

Do not install AI/ML libraries into the system Python. Every major framework pins transitive dependencies aggressively and will break your distro packages. Use isolated virtual environments per project, or conda/mamba when you need compiled CUDA kernels managed for you.

Option A — venv (lightweight, always available)

python3 -m venv ~/.venvs/ailab
source ~/.venvs/ailab/bin/activate
pip install --upgrade pip setuptools wheel

Option B — Miniforge (recommended for GPU work)

curl -fsSLo miniforge.sh \
  https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.sh
bash miniforge.sh -b -p "$HOME/miniforge3"
"$HOME/miniforge3/bin/conda" init bash
source ~/.bashrc
conda create -n ailab python=3.11 -y
conda activate ailab

NVIDIA CUDA Driver Setup

The proprietary driver is the single most common stumbling block. On all three families the procedure is: disable Nouveau, install the driver package, then verify with nvidia-smi.

Debian / Ubuntu

sudo apt update
sudo apt install -y linux-headers-$(uname -r) software-properties-common
sudo add-apt-repository contrib   # Debian only
sudo apt install -y nvidia-driver firmware-misc-nonfree   # Debian
# Ubuntu: ubuntu-drivers autoinstall  OR  apt install nvidia-driver-545

Fedora / RHEL / Rocky

# Add RPM Fusion
sudo dnf install -y \
  https://download1.rpmfusion.org/free/fedora/rpmfusion-free-release-$(rpm -E %fedora).noarch.rpm \
  https://download1.rpmfusion.org/nonfree/fedora/rpmfusion-nonfree-release-$(rpm -E %fedora).noarch.rpm
sudo dnf install -y akmod-nvidia xorg-x11-drv-nvidia-cuda
sudo systemctl reboot

Arch

sudo pacman -S --needed nvidia nvidia-utils cuda cudnn
sudo systemctl reboot

After reboot, verify the driver loaded and check compute capability:

nvidia-smi
nvcc --version   # if cuda toolkit is installed

Modern ML Stack: PyTorch and JAX

PyTorch (CUDA)

Always install from the official index URL for the correct CUDA version. Mixing conda and pip here is safe as long as you install PyTorch last.

pip install torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/cu121

Quick smoke test:

python3 -c "
import torch
print(torch.__version__)
print('CUDA available:', torch.cuda.is_available())
print('Device:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'CPU')
"

JAX (NVIDIA)

pip install -U "jax[cuda12]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

python3 -c "import jax; print(jax.devices())"

ROCm for AMD GPUs

ROCm 6.x supports RDNA2 and newer plus most CDNA cards. Install the ROCm base stack via the official AMDGPU installer, then install PyTorch's ROCm wheel:

# After amdgpu-install --usecase=rocm from repo.radeon.com:
pip install torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/rocm6.0

Classic Genetic Programming and Evolutionary Frameworks

DEAP — Distributed Evolutionary Algorithms in Python

DEAP is the go-to library for genetic programming, evolution strategies, and multi-objective optimization (NSGA-II, SPEA2). It is pure Python and installs trivially.

pip install deap

A minimal evolving symbolic regression in one shot:

python3 - <<'EOF'
import operator, math, random
from deap import algorithms, base, creator, gp, tools

pset = gp.PrimitiveSet("MAIN", 1)
pset.addPrimitive(operator.add, 2)
pset.addPrimitive(operator.mul, 2)
pset.addPrimitive(math.sin, 1)
pset.renameArguments(ARG0="x")

creator.create("FitnessMin", base.Fitness, weights=(-1.0,))
creator.create("Individual", gp.PrimitiveTree, fitness=creator.FitnessMin)

toolbox = base.Toolbox()
toolbox.register("expr", gp.genHalfAndHalf, pset=pset, min_=1, max_=3)
toolbox.register("individual", tools.initIterate, creator.Individual, toolbox.expr)
toolbox.register("population", tools.initRepeat, list, toolbox.individual)

def evaluate(individual):
    func = toolbox.compile(expr=individual)
    return (sum((func(x) - x**2)**2 for x in range(-5, 6)),)

toolbox.register("compile", gp.compile, pset=pset)
toolbox.register("evaluate", evaluate)
toolbox.register("select", tools.selTournament, tournsize=3)
toolbox.register("mate", gp.cxOnePoint)
toolbox.register("mutate", gp.mutUniform, expr=toolbox.expr, pset=pset)

pop = toolbox.population(n=100)
hof = tools.HallOfFame(1)
algorithms.eaSimple(pop, toolbox, 0.7, 0.1, 40, halloffame=hof, verbose=False)
print("Best:", hof[0])
EOF

PyGAD — Genetic Algorithms with NumPy/PyTorch Integration

pip install pygad

PyGAD has first-class support for evolving PyTorch and Keras model weights, making it practical for neuroevolution tasks without a full ES library.

Nevergrad (Meta) — Gradient-Free Optimization

pip install nevergrad

Nevergrad bundles CMA-ES, DE, PSO, and dozens of other algorithms behind a single consistent API. Useful when your fitness landscape is non-differentiable.

Artificial-Life Simulators

Avida-ED and Avida (CLI)

Avida is the canonical digital-organism evolution platform from the Pennock/Lenski group. Build from source on Debian/Ubuntu:

sudo apt install -y build-essential cmake git
git clone https://github.com/devosoft/avida.git
cd avida
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)
./avida -c ../support/avida.cfg

Xentica — GPU-Accelerated Cellular Automata (Python)

pip install xentica

Xentica uses PyCUDA to run large-scale CA experiments on GPU. It requires a working CUDA install. Check their documentation for the exact PyCUDA version pin before installing.

NetLogo via OpenJDK (Agent-Based Modeling)

NetLogo is the standard for agent-based a-life models and runs fine under Linux with OpenJDK 17+.

# Debian/Ubuntu
sudo apt install -y openjdk-21-jre
wget https://ccl.northwestern.edu/netlogo/6.4.0/NetLogo-6.4.0-64.tgz
tar xzf NetLogo-6.4.0-64.tgz
cd NetLogo-6.4.0-64
./NetLogo   # launches GUI; works on Wayland via XWayland

Wayland note: NetLogo's Java Swing UI runs through XWayland automatically on modern desktops. No extra flag needed on GNOME 45+ or KDE Plasma 6.

Reinforcement Learning Environments

pip install gymnasium[all] stable-baselines3 shimmy

Gymnasium (the maintained OpenAI Gym fork) plus Stable-Baselines3 gives you a complete RL sandbox. For MuJoCo physics, install the free mujoco Python package directly — the old licensing restriction is gone:

pip install mujoco

Monitoring GPU Utilization

Long-running training sessions need live resource visibility. nvtop is the best interactive monitor and handles both NVIDIA and AMD cards:

# Debian/Ubuntu
sudo apt install nvtop

# Fedora
sudo dnf install nvtop

# Arch
sudo pacman -S nvtop

nvtop

For headless servers, log GPU stats every 5 seconds with:

nvidia-smi dmon -s u -d 5

Troubleshooting

• CUDA out-of-memory on startup: Another process holds the GPU. Run nvidia-smi to find the PID and kill it, or set PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True in your environment.
• Wrong CUDA version mismatch: torch.version.cuda must match the driver's supported CUDA. Upgrade the driver or downgrade the PyTorch wheel index URL.
• Avida build fails on GCC 13+: Pass -DCMAKE_CXX_FLAGS="-Wno-error" to cmake; some deprecated C++ constructs trigger errors by default on modern GCC.
• NetLogo blank window on Wayland: Force XWayland explicitly with GDK_BACKEND=x11 ./NetLogo if auto-detection fails.
• JAX sees only CPU despite GPU present: Confirm libcuda.so.1 is on LD_LIBRARY_PATH. With Miniforge: conda install -c nvidia cuda-nvcc often resolves it.