New metabolic multi-agent optimizer MMAO-Dyn beats PSO and DE on dynamic problems
This bio-inspired optimizer reallocates resources to recover from environmental changes, outperforming classic swarm and evolutionary methods.
The Metabolic Multi-Agent Optimizer (MMAO) draws inspiration from biological metabolism, using agents that consume and share energy to find optima in static landscapes. But real-world problems often change over time—factories retool, markets shift, sensor drifts occur. In a new paper, Jinliang Xu and Liping Ma extend MMAO to handle these dynamic environments with MMAO-Dyn. The key innovation is keeping the core metabolic control loop intact: agents still manage private energy, communal budgets, role drift, success feedback, and lifecycle turnover, but now these mechanisms are mapped to a nonstationary setting where previously useful patterns become obsolete. This avoids replacing the core loop with external adaptation modules, preserving the biological analogy while gaining dynamic robustness.
Evaluated across 18 scenarios (shifted sphere, Ackley, Rastrigin at 10D, 20D, 30D with two change severities and 12 seeds per scenario), MMAO-Dyn achieved a mean offline error of 28.07 over 216 runs—significantly better than Dynamic PSO-lite (34.65), Dynamic DE-lite (67.09), and Dynamic Random Search (111.37). The gains are clearest in post-change recovery speed, especially on sphere and Rastrigin families. However, the authors note that the win-loss count against Dynamic PSO-lite remains unfavorable, and the NoMemoryRefresh ablation stays very close to the full method. Thus, MMAO-Dyn is positioned as a credible family expansion for MMAO: its main value lies in recovery-oriented resource redistribution rather than universal dominance, offering a fresh bio-inspired approach for dynamic optimization problems across engineering and AI domains.
- MMAO-Dyn maps metabolic concepts (private energy, communal budget, role drift, success feedback, lifecycle turnover) directly to dynamic optimization without external adaptation modules.
- Achieves mean offline error 28.07 across 216 runs on 10D–30D shifted landscapes, outperforming Dynamic PSO-lite (34.65) and Dynamic DE-lite (67.09).
- Strongest in post-change recovery speed; the NoMemoryRefresh ablation remains close, showing recovery-oriented resource redistribution is the core strength.
Why It Matters
A bio-inspired optimizer that efficiently reallocates resources in changing environments, useful for real-world optimization in robotics, finance, and engineering.