metaguided

Metaguided is an adjective for systems, methods, or processes that rely on meta-information to steer decisions or search trajectories. The guiding information comes from levels above the immediate task, such as historical performance, task descriptors, or learned meta-strategies. In practice, metaguided approaches use this meta-information to select methods, adjust parameters, or switch strategies during execution. The term is not universally standardized and is often used to describe meta-level control.

Mechanisms include meta-learning, where a controller optimizes future learning, and meta-heuristics, which apply high-level rules to guide search. Adaptive controllers and auto-tuning systems reconfigure hyperparameters based on observed performance. Across these implementations, meta-information may be performance histories, feature descriptors, or reliability estimates of subcomponents.

Applications appear in artificial intelligence, optimization, robotics, and game design. For example, a metaguided reinforcement learning agent might choose among sub-policies using a meta-controller. An optimization routine might adjust exploration rate or switch search strategies according to prior iterations. In games, metaguided techniques can adapt strategy selection to shifts in the evolving metagame.

As a relatively informal label, metaguided concepts overlap with meta-learning, hyperparameter optimization, meta-heuristics, and adaptive systems. Challenges include added computational overhead, potential overfitting to meta-features, and the opacity of meta-guiding signals.