Package-level declarations
Types
A write that failed partway through applying a plan.
How a rule searches ActionSet for its best member. Model-independent, which is why it belongs to the library: the loop is the same for every value-function and cost-function policy ever written, and only the scoring differs.
An action that could not be applied: outside a lever's bounds, non-integral on an integer domain, NaN, the wrong arity, or violating a joint constraint.
Thrown by leverFor/rewardFor when the owner backs more than one declared lever or term.
A declared name that does not resolve against the element it was given to.
§4.3.1 — name resolution for one element's surface. Built only by DecisionElementBuilder.
Everything a rule may know at a decision instant.
§4.10.1 and §4.10.2. Declared through decisionElement; never constructed directly.
Score every action. Requires a finite ActionSet.size.
A constant action. Arity is checked against the declaration at assignment.
Score a regular grid over each lever's feasible range, keeping only feasible points. The fallback when the set is continuous or too large to walk.
A value function that improves from observed experience.
The descriptive half of a lever. Holding one cannot write anything (§4.3.1.1).
The skeleton shared by every policy that chooses among available actions rather than constructing one directly: value-function approximations, and cost-function approximations that score candidates.
A policy with a lifecycle: per-replication state to reset, or a resource to release.
An attempt to narrow a lever outside what the model declares, or to non-integral bounds on an integer domain.
Do nothing — every lever takes its declared neutral value. The Level-2 compatibility baseline of §6.2.
A form the DSL accepts and the library does not yet carry (G.9 row 11).
A decision element reviewed on a fixed period, declared and scheduled in one construction.
Thrown when a policy schedules an event during the call that asked it for an action.
S§C.11.1 — thrown when decide is entered while a decision is already in progress.
A reward term whose declared accumulation kind does not match what its source actually offers — a TIME_INTEGRAL declared against a plain Response, say.
Read-only view of a Response, TWResponse, or Counter used as a reward source.
Moved here from the implementations package, because it is not one.
D11 — thrown when deferred requests keep arriving during the decisions that answer them.
What produced a set of transitions: which model, which experiment, which element, which rule — and the descriptor that gives the rows their meaning.
Score candidates actions drawn at random from the set.
Implement when a rule requires something of the declared shape, or must BUILD something from it — a discretization grid, a table sized to the action vector, a network's input layer.
Thrown when a DecisionContext is read outside the action call that supplied it.
§4.8.3. The atom of recorded experience.
§4.8.2. A write-only consumer with a per-experiment lifetime.
How much of a decision surface's units were declared, and how much of the checking that makes possible actually ran (§4.2.4, G.9 row 7).
An approximation of downstream value, evaluated at a post-decision state — the V̄ of approximate dynamic programming (§4.5.5).
Properties
Functions
Entry point, shaped like ModelElement.queueingNetwork (§4.1.2.1).