---

🌟 AutoPSO: A Metaframework for Automated Particle Swarm Optimization 🌟

---

AutoPSO is an online framework for automatically constructing PSO variants. Traditional PSO research has produced many useful mechanisms, such as different parameter schedules, exemplar choices, and swarm topologies. In practice, however, turning those mechanisms into a strong optimizer often still depends on manual trial-and-error: a researcher selects components, tunes coefficients, runs a benchmark, then repeats the process for the next problem.

AutoPSO turns this hand-crafted design loop into a black-box optimization problem. An outer PSO searches over a compact configuration space of PSO components, while many inner generalized PSO solvers are instantiated in parallel to solve the target problem and report feedback. With EvoX and PyTorch, those inner runs are tensorized and mapped naturally to modern GPUs, making it practical to adapt PSO designs online under a fixed wall-clock budget.

Method

Each outer particle encodes one candidate PSO variant:

• weight: the inertia coefficient shared by two update policies.
• k1, k2, k3, k4: acceleration coefficients for the two policies.
• type1, type2, type3, type4: exemplar choices selected from a nine-candidate pool.
• percent: the population split ratio for two sub-swarms.

The inner solver is a generalized PSO. At every step, it partitions the swarm into two groups, assigns each group a different update policy, and selects exemplars from a shared pool including current position, personal best, global best, population center, random individuals, random personal bests, random elites, and newly sampled candidates.

Main Result

The paper evaluates AutoPSO against six classical PSO variants on CEC2022 under equal wall-clock time. For 20D functions, each run is limited to 120 seconds and reported curves are averaged over 31 independent runs. The main CEC2022 setting implemented in this repository uses:

outer optimizer: PSO
outer population: 100
inner optimizer: AutoPSO / generalized PSO
inner population: 100
inner iterations per outer evaluation: 800
time budget: 60s for CEC2022 10D, 120s for CEC2022 20D

Reproduction note: the paper reports wall-clock experiments on an RTX 3090 GPU. If PyTorch is running on CPU, the same 120-second budget will perform far fewer outer iterations and the results may be noticeably worse. Check your backend with python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))".

Repository Layout

assets/                         Logos and paper figures used by the README
examples/pytorch/example_cec2022.py  PyTorch CEC2022 example entry point
scripts/summarize_json_results.py Result summarizer for JSON outputs
src/autopso/                     AutoPSO package source
show_env.py                      Environment and dependency checker

Installation

Create or activate an environment with PyTorch and EvoX. For CPU-only smoke tests:

pip install -e ".[pytorch]"

For GPU examples, install the CUDA-enabled PyTorch build that matches your CUDA driver, then install EvoX. A typical setup is:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install -U "evox[default]"
pip install -e ".[pytorch]"

Verify the environment:

python show_env.py
python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))"

Quick Start

Run the CEC2022 test:

python examples/pytorch/example_cec2022.py \
  --dims 20 \
  --functions 1 2 3 4 5 6 7 8 9 10 11 12 \
  --runs 31 \
  --max-time 120 \
  --device cuda \
  --output outputs/cec2022_20d_31runs_pytorch

Summarize results:

python scripts/summarize_json_results.py outputs/cec2022_20d_31runs_pytorch/autopso_cec2022_20D.json

The runner prints progress during each run. Use --progress-interval 1 for more frequent updates, or --progress-interval 0 to disable progress output.

The PyTorch runner uses the AutoPSO encoding and CEC2022 data packaged in this repository. Its CEC2022 problem class follows the EvoX PyTorch Problem style when evox.core is available, with a fallback for older EvoX installs.

Citation

If you use this code, please cite the corresponding paper:

@article{yu2026autopso,
  title   = {AutoPSO: A Meta-Framework for Automated Particle Swarm Optimization},
  author  = {Yu, Xinmeng and Gao, Jiaxin and Zhang, Jianguo and Jiang, Dongmei and Cheng, Ran},
  journal = {IEEE Transactions on Evolutionary Computation},
  year    = {2026},
  doi     = {10.1109/TEVC.2026.3718908}
}

Acknowledgements

AutoPSO is implemented with the EvoX framework for vectorized evolutionary computation and hardware acceleration.