RL-SOMA: Reinforcement learning control of the Self-Organizing Migrating Algorithm for global optimization and UAV path planning

The Self-Organizing Migrating Algorithm (SOMA) moves candidate solutions along discretized trajectories towards leader individuals, under two sensitive control parameters and a migration strategy fixed before the run. No single setting suits every landscape or every phase of a run. We propose RL-SOMA, in which a tabular Q-learning controller picks, at every generation, an operator and its parameter preset from a memetic portfolio of six operators, conditioned on a discretized state of population diversity, stagnation and budget progress, and warm-started by a progress-aware informed prior so that learning refines a competent schedule within one run. On the classical 13-function suite at D = 30 RL-SOMA leads the SOMA family, winning at least 10 of 13 functions against every SOMA variant, and ranks third of the eight methods behind GWO and L-SHADE, with all three inside the critical difference. Translating the optima of the nine centre-optimum functions under five independent offsets shows GWO’s lead to be geometric: its median result degrades by 10.6 orders of magnitude, while RL-SOMA is unchanged. On CEC2017 at D = 30 RL-SOMA ranks 3.72, indistinguishable from SOMA-T3A and DE, with L-SHADE ahead of the field. On the classical suite at D = 50 it ranks second (2.62), behind GWO (1.96) and ahead of L-SHADE (3.04). On constrained three-dimensional UAV path planning over five spline-parameterized scenarios RL-SOMA attains the best average rank among seven planners, L-SHADE included (1.60), is never significantly outperformed on any scenario, and holds 100% feasibility on every map under both the collision-free and the stricter margin criterion, where PSO falls to 55% and GWO to 80%. An ablation with a frozen-prior baseline separates the two learning components: the prior carries most of the gain, while online learning adds a significant further improvement on three of six probes and degrades none.

Authors

Institutions

Publication Details

Journal
Engineering Science and Technology an International Journal
Published
2026-11-13
DOI
https://doi.org/10.1016/j.jestch.2026.102590
Primary Topic
Metaheuristic Optimization Algorithms Research
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

RL-SOMA: Reinforcement learning control of the Self-Organizing Migrating Algorithm for global optimization and UAV path planning

Hoang Thi Minh Anh, Quoc Bao Diep, Minh Tam T. Phan, Minh Diep
1 links
Engineering Science and Technology an International Journal
Metaheuristic Optimization Algorithms Research
article

RL-SOMA: Reinforcement learning control of the Self-Organizing Migrating Algorithm for global optimization and UAV path planning

Hoang Thi Minh Anh, Quoc Bao Diep, Minh Tam T. Phan, Minh Diep
article en

Abstract

The Self-Organizing Migrating Algorithm (SOMA) moves candidate solutions along discretized trajectories towards leader individuals, under two sensitive control parameters and a migration strategy fixed before the run. No single setting suits every landscape or every phase of a run. We propose RL-SOMA, in which a tabular Q-learning controller picks, at every generation, an operator and its parameter preset from a memetic portfolio of six operators, conditioned on a discretized state of population diversity, stagnation and budget progress, and warm-started by a progress-aware informed prior so that learning refines a competent schedule within one run. On the classical 13-function suite at D = 30 RL-SOMA leads the SOMA family, winning at least 10 of 13 functions against every SOMA variant, and ranks third of the eight methods behind GWO and L-SHADE, with all three inside the critical difference. Translating the optima of the nine centre-optimum functions under five independent offsets shows GWO’s lead to be geometric: its median result degrades by 10.6 orders of magnitude, while RL-SOMA is unchanged. On CEC2017 at D = 30 RL-SOMA ranks 3.72, indistinguishable from SOMA-T3A and DE, with L-SHADE ahead of the field. On the classical suite at D = 50 it ranks second (2.62), behind GWO (1.96) and ahead of L-SHADE (3.04). On constrained three-dimensional UAV path planning over five spline-parameterized scenarios RL-SOMA attains the best average rank among seven planners, L-SHADE included (1.60), is never significantly outperformed on any scenario, and holds 100% feasibility on every map under both the collision-free and the stricter margin criterion, where PSO falls to 55% and GWO to 80%. An ablation with a frozen-prior baseline separates the two learning components: the prior carries most of the gain, while online learning adds a significant further improvement on three of six probes and degrades none.

Engineering Science and Technology an International JournalVol. 84
Vietnam Posts and Telecommunications Group (Vietnam) (VN), Ho Chi Minh City University of Industry and Trade (VN)
Metaheuristic Optimization Algorithms Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon