- Sequential learning learns which actions lead to better delayed rewards.
- Transition futures learn what measured outcomes can follow an action.
- Probabilistic planning compares those outcomes against the declared objective.
sequential_q_mlp into a distributional model,
nor does it remove the sequential learner’s validation requirements.
Here is a complete combined Domain declaration. It uses the same transition
task as the single-step planner example and adds episode reward learning:
relation: "controls", the pre-action context, and request
decision_id, transition_plan and selection. Choose deliberately whether
your application follows the planner or the contextual selector. Keep
policy_weight: 0 initially to test transition planning independently; a positive
value adds the policy learner’s first-action utility preference.
For a planner-selected steady action, send this to /feedback after execution,
replacing the decision ID, episode information and observations with real values:
terminal: true only when the episode actually
ends. next_state is the real observed next context; x remains the pre-action
input. All action names, relation names and metadata paths must match the Domain.
If the query used a future_id, include it in the feedback metadata too.
Inspect transition_eligibility.futures_update for forecast resolution and
learning_state.subsystems.feedback_policy for policy learning. The two counters
measure different observations and need not match. Do not post the same execution
again to /events as an extra sample.