Adaptive Ant Colony Optimization in Dynamic Path Planning

Abstract

Dynamic path planning—finding safe, efficient routes while the environment changes—remains a core challenge in autonomous robotics, intelligent transportation, and logistics. Classical Ant Colony Optimization (ACO) is attractive for path planning due to its distributed search, positive feedback, and robustness to local optima; however, it degrades when costs, obstacles, or constraints change during execution because its pheromone model encodes stale information. This manuscript proposes an Adaptive Ant Colony Optimization (A-ACO) framework tailored for dynamic environments. The framework introduces four complementary mechanisms: (i) event-triggered pheromone aging that increases evaporation locally and temporarily after detected changes; (ii) memory-aware partial reinitialization that resets pheromone in neighborhoods of change while preserving global structure; (iii) online heuristic shaping using short-horizon obstacle forecasts to bias ants away from emergent hazards; and (iv) time-bounded anytime re-optimization that reuses incumbent solutions under iteration budgets for real-time response. We formalize the transition rule, pheromone update, and change-aware schedules, and we provide complexity insights.

A simulation campaign on 2D occupancy grids with moving obstacles compares A-ACO against Adaptive Candidate System (ACS/ACO baseline), D* Lite, and RRT*, under three dynamics levels (low/medium/high). Across 90 runs per algorithm, A-ACO reduces average path cost by 6–13%, decreases replanning latency by 18–35%, and improves success rate by 2–7 percentage points relative to baselines, while maintaining collision rates near zero. A two-way mixed ANOVA (factor: algorithm; repeated factor: dynamics level) shows a significant main effect of algorithm on path cost and latency (p < .001), with A-ACO outperforming all comparators in post-hoc tests. The results suggest that local, event-aware pheromone management and predictive heuristic shaping are decisive for dynamic feasibility and responsiveness. We conclude with limitations (sensor noise, non-holonomic kinematics) and future directions (multi-robot coordination, risk-aware multiobjective extensions).