DeterministicPolicy<T>

Models & Types · AiDotNet.ReinforcementLearning.Policies

Deterministic policy for continuous action spaces. Directly outputs actions without sampling from a distribution. Commonly used in DDPG, TD3, and other deterministic policy gradient methods.

Inheritance: ModelBase<T, Vector<T>, Vector<T>> → PolicyBase<T> → DeterministicPolicy<T>

Implements: IGeneratedParameterRegistrar<T>

Example

// Create a deterministic policy for DDPG/TD3 continuous control
var network = new NeuralNetwork<double>();
var policy = new DeterministicPolicy<double>(network, actionSize: 2, new EpsilonGreedyExploration<double>());
            
// Get a deterministic action with exploration noise
var state = new Vector<double>(new double[] { 0.5, -0.3, 1.0, 0.2 });
var action = policy.SelectAction(state, training: true);

Constructors

Constructor Summary
DeterministicPolicy() Initializes a new instance with default settings.
DeterministicPolicy(DeterministicPolicyOptions<T> options) Builds the policy from options: a dense ReLU network of the options’ hidden widths from StateSize inputs to ActionSize outputs, explored with the options’ strategy or, when it is null, Ornstein-Uhlenbeck noise (θ = 0.15, σ = 0.2) sized to…
DeterministicPolicy(INeuralNetwork<T> policyNetwork, DeterministicPolicyOptions<T> options) Initializes a new instance of the DeterministicPolicy class from its options: action size, exploration strategy, tanh squashing and Seed.
DeterministicPolicy(INeuralNetwork<T> policyNetwork, int actionSize, IExplorationStrategy<T> explorationStrategy, bool useTanhSquashing, Random random) Initializes a new instance of the DeterministicPolicy class.

Methods

Method Summary
IReadOnlyList<INeuralNetwork<T>> GetNetworks() Gets the neural networks used by this policy.
T ComputeLogProb(Vector<T> state, Vector<T> action) Computes log probability for a deterministic policy.
Vector<T> SelectAction(Vector<T> state, bool training) Selects a deterministic action from the policy network.
void Reset() Resets the exploration strategy.