Overview
MeEvo is an Automatic Heuristic Design (AHD) framework that integrates Metacognitive Evolution with Natural Evolution. This framework aims to enhance heuristic generation through a cyclical process, addressing identified limitations in current Large Language Model (LLM)-based AHD architectures.
Research Context
Large Language Models (LLMs) have advanced Automatic Heuristic Design (AHD) by facilitating heuristic generation via reasoning and code synthesis. In this paradigm, an LLM formulates an algorithm design and then synthesizes executable heuristic code based on that design.
Existing LLM-based AHD architectures typically fall into two main categories:
- Natural Evolution: This approach applies genetic operators, specifically crossover and mutation, to the generated heuristic code to explore a diverse range of strategies. However, a limitation of Natural Evolution is its tendency to discard the underlying reasoning traces that informed the design decisions, which weakens knowledge retention.
- Metacognitive Evolution: In contrast, Metacognitive Evolution retains these reasoning traces and refines them through a process of reflection. A key limitation here is the absence of population-level recombination, which restricts the scope of exploration.
These distinct limitations in existing architectures, namely weakened knowledge retention in Natural Evolution and restricted exploration in Metacognitive Evolution, collectively contribute to reduced search efficiency, diminished stability, and lower solution quality when applied to complex problems.
Approach
MeEvo addresses the limitations of existing LLM-based AHD architectures by cyclically combining Natural Evolution and Metacognitive Evolution. The framework incorporates an operator balance mechanism designed to shift from exploration to exploitation throughout its operation.
The operational cycle of MeEvo involves two interconnected phases:
- Natural Evolution Phase: During this phase, Natural Evolution explores heuristic code. As it does so, it records several key pieces of information into a shared history. This recorded history includes LLM-generated reasoning traces, computed fitness values, encountered errors, and the best heuristic identified thus far.
- Metacognitive Evolution Phase: Following the Natural Evolution phase, Metacognitive Evolution processes the shared history. It reflects upon this historical data to generate improved heuristics. These enhanced heuristics are then fed back into the subsequent cycle of Natural Evolution.
This integrated design ensures that population-driven exploration, characteristic of Natural Evolution, and reflection-driven refinement, central to Metacognitive Evolution, mutually reinforce each other within the MeEvo framework.
Findings
Experiments were conducted on five distinct optimization problems to evaluate MeEvo's performance against tested LLM-based AHD architectures.
- MeEvo demonstrated stronger overall performance.
- MeEvo exhibited lower variance in its results compared to the other architectures.
- The performance advantages of MeEvo were particularly notable on complex constrained tasks.
Why This Matters
The observed stronger performance and lower variance of MeEvo, especially on complex constrained tasks, indicate a potential for more robust and effective automatic heuristic design. This addresses existing limitations in search efficiency, stability, and solution quality identified in current LLM-based AHD architectures.