Code and environment instances for: Neural-Network-Guided Particle Swarm Optimization for Tracking Moving Optima in Continuous Dynamic Environments

Published: 2 November 2026| Version 1 | DOI: 10.17632/gm82kx49db.1
Contributors:
,

Description

Code and environment instances supporting the article "Neural-Network-Guided Particle Swarm Optimization for Tracking Moving Optima in Continuous Dynamic Environments" (S. Raharja and T. Sugawara, Swarm and Evolutionary Computation, 2026). The code folders (1, 2, 3, 6, 7) are zip archives because of the repository's folder-depth limit; the environment folders (4, 5) are not zipped. 1_simulator: Java 21 simulator running CNNPSO, DNNPSO and the PSO, SPSO, PSPSO and DMPSO baselines. 2_orchestrator: Java 21 orchestrator for environment generation, the auxiliary-PSO supervisory signal a(t), network pre-training and run configurations; configs/ holds the configurations used in the article. 3_analyser: Python package computing the tracking metrics and the environment-level statistics. 4_environments_pre_fix: the 60 environment instances (30 main, 30 destination-density) as originally generated, where the stored reference optimum is the auxiliary canonical PSO's estimate, the signal the networks were trained on. JSON (zstd-compressed) and compact npz. 5_environments_certified: the 30 main instances with the certified grid reference optimum a'(t) written in, against which every reported result is measured, plus the per-tick certified reference for all 60 instances. 6_gleet_patch: our modifications to GLEET (Ma et al., GECCO 2024) as a patch against upstream commit e7159273, with the evaluated checkpoint. GLEET itself is not redistributed. 7_reference_solver: the certified reference-optimum solver, the rescoring script and the scripts producing the published table cells and Figures 5-6. README.md documents usage, file formats, provenance and checksums. Each instance stores the full trajectory of every peak, so the evaluation set is reproduced exactly. Training, the auxiliary PSO and the swarms are stochastic, so re-running the experiments reproduces the reported statistics in distribution.

Files

Steps to reproduce

0. Unzip the code archives in place: 1_simulator.zip, 2_orchestrator.zip, 3_analyser.zip, 6_gleet_patch.zip, 7_reference_solver.zip. 1. Decompress the environments: zstd -d --long=27 env_k.json.zst, then check with JSON_SHA256SUMS.txt. 2. Build the simulator and orchestrator (Java 21, Maven): mvn clean package in 1_simulator and 2_orchestrator. 3. Run on the shipped instances: java -jar orchestrator.jar -c configs/config.json -o <out> -u <decompressed journal_30> -x <simulator jar>. For the destination-density set use configs/config_g3_env_20260817.json with g3_30. 4. The certified reference is supplied in 5_environments_certified; to recompute it: python 7_reference_solver/certify_aopt.py <npz_dir> <out_dir> 0 1 ... 29. 5. Rescore every run against it with SWEVO_ROOT=<output root> python 7_reference_solver/rescore_runs_v3.py main (and g3). Then run the analyser's round2-rescore and round2-stats commands (3_analyser/README.md), r3_v3adapt.py and r3_build.py for the table cells, and r3_figs.py for Figures 5-6. 6. GLEET: clone MetaEvo/GLEET, check out e7159273d4dc7ea78ef3ae3a701069cfd4d61191, git apply 6_gleet_patch/dnnpso_gleet.patch, and evaluate with eval_dnnpso_gleet.py and trained_checkpoint/latest.pt, one process per environment. Full details are in README.md.

Institutions

Categories

Artificial Neural Network, Evolutionary Computation, Particle Swarm Optimization, Swarm Intelligence Algorithm

Funders

Licence