Interface TrainingRules<State, Command, Event>

interface TrainingRules<State, Command, Event> {
    observationVersion: string;
    rule: SimulationRule<State, Command, Event, Random.Generator>;
    command(state: State, actions: readonly (number | null)[]): Command;
    finished?(state: State): boolean;
    initial(random: Random.Generator): State;
    observe(state: State): readonly NeuralObservation[];
    rewards(
        before: State,
        command: Command,
        outcome: SimulationStep<State, Event>,
    ): readonly number[];
}

Type Parameters

State
Command
Event

Index

Properties

observationVersion rule

Methods

command finished? initial observe rewards

Properties

observationVersion

observationVersion: string

rule

rule: SimulationRule<State, Command, Event, Random.Generator>

Same rule used by the playable game; a command may advance a turn or fixed ticks.

Methods

command

command(state: State, actions: readonly (number | null)[]): Command
Parameters
- state: State
- actions: readonly (number | null)[]
Returns Command
- Defined in src/simulation/Training.ts:15

`Optional`finished

finished?(state: State): boolean
Detect an already-terminal reset, without taking a dummy action.
Parameters
- state: State
Returns boolean
- Defined in src/simulation/Training.ts:19

initial

initial(random: Random.Generator): State
Parameters
- random: Random.Generator
Returns State
- Defined in src/simulation/Training.ts:11

observe

observe(state: State): readonly NeuralObservation[]
Parameters
- state: State
Returns readonly NeuralObservation[]
- Defined in src/simulation/Training.ts:14

rewards

rewards(
    before: State,
    command: Command,
    outcome: SimulationStep<State, Event>,
): readonly number[]
One reward per agent, in the observation's stable roster order.
Parameters
- before: State
- command: Command
- outcome: SimulationStep<State, Event>
Returns readonly number[]
- Defined in src/simulation/Training.ts:17