Original title: Towards Autonomous Bio-Inspired Optimization: Deep Reinforcement Learning for Adaptive Metaheuristic Orchestration
César Carrasco, Broderick Crawford, Carlos Valle, Roberto Zulantay +2 · September 2026
Solving complex matching or packing problems is difficult because no single algorithm works best all the time. Researchers wanted to see if an artificial intelligence could swap between different nature-inspired algorithms while working on the same problem. They used deep reinforcement learning, a method where an agent learns through trial and error rewards. The agent managed seven different algorithms that shared one group of evolving solutions. Every fixed number of steps, the agent checked the search progress and picked which algorithm took over next. The researchers found that this selection agent beat simple random choices and switching in turns. When tested on a set covering problem, it placed first out of eleven strategies. At half the testing budget on a knapsack problem, the agent methods took the top two spots. Testing revealed the system mainly succeeds by learning which weak algorithms to avoid entirely.
You have a team of chefs with different cooking styles. Each style works well for some dishes, but not others. The problem is you don't know which chef is best for the next step without trying them, and switching chefs wastes the prep work already done. In this paper, the chefs are different search methods for solving hard problems, and the recipe is the plan for finding the best solution. The agent learns, over time, which chefs to avoid for certain tasks, handing the partially prepared dish to the next chef.
Choosing the right optimization method is usually a trial-and-error process. An automated agent that learns to pick and skip methods can adapt to different problems, achieving high-quality solutions with less human effort. This is especially useful when no single method consistently outperforms others, as shown on set covering problems.
The joint selection-and-configuration agent did not improve over blind baselines, suggesting that learning both which metaheuristic and how to configure it is harder. The policy mainly learns which methods to avoid rather than which to actively choose. Results are limited to two binary combinatorial optimization problems (MKP and SCP) with specific benchmark instances, so generalization to other problem types is not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Towards Autonomous Bio-Inspired Optimization: Deep Reinforcement Learning for Adaptive Metaheuristic Orchestration — César Carrasco, Broderick Crawford, Carlos Valle et al. — Biomimetics — doi:10.3390/biomimetics11090682
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.