How learning and planning work together
The application presents the current observation and available actions. The planner compares their possible continuations, and the executor carries out the selected action or sequence. The application then observes what actually happened. Eligible observations and outcomes update the relevant learner. Transition learning improves predictions of action effects; value learning revises estimates of useful outcomes. Procedural acquisition and refinement produce reusable learned action sequences for later selection. Planning influences learning through the actions selected, while learning influences later planning through retained predictions, values, or skills. Search-derived assessments can also contribute to training targets. At each recorded state, define how the planning assessment and observed outcome form the target, and preserve the provenance of both components.What a plan compares
A planning task needs a current state, available actions, a way to obtain successors, and a way to evaluate them. These parts can be combined in several ways:- Exact transitions and learned values: task rules determine what happens next, while a learned evaluator estimates unfinished states.
- Learned transitions and a declared goal: a transition model predicts successive states, and the goal supplies their utility.
- Learned transitions and learned values: one model predicts action effects and another scores the resulting states.
- Retained action sequences: each candidate is a complete acquired or refined skill, evaluated according to its outcome.
Choose the next step
- Learn and refine state values: fit on outcomes, reconsider legal alternatives, learn grouped differences, continue from retained parameters, and adopt a candidate after comparison.
- Value API reference: use the eight value routes and inspect model, candidate, and training state.
- Native planning options: choose evaluator source, depth, work limits, extensions, caching, and exploration.
- Planning integration: connect actions, successors, and evaluators through HTTP or streaming search, then inspect the returned diagnostics.
- Planning over learned transitions: configure Domain-query planning, state bindings, goals, and horizons.
- Combine predictions and values: connect separate transition and value learners while tracking their identities and update behavior.
- Decisions beyond the next move: distinguish planning horizon, learning credit, and the execution boundary.
- Select from retained sequences: bind candidates to executable commands and preserve complete-sequence commitment.
- Interactions during frozen use: keep every participating learner fixed when evaluating retained behavior.
