A3CAgent<T>

Models & Types · AiDotNet.ReinforcementLearning.Agents.A3C

Asynchronous Advantage Actor-Critic (A3C) agent for reinforcement learning.

Inheritance: ReinforcementLearningAgentBase<T> → DeepReinforcementLearningAgentBase<T> → A3CAgent<T>

Implements: IGradientComputable<T, Vector<T>, Vector<T>>

For Beginners

A3C is like having multiple students learn simultaneously - each has different experiences, and they periodically share knowledge with a “master” network. This parallel learning provides stability and diverse exploration.

Key features:

  • Asynchronous Updates: Multiple workers update global network independently
  • No Replay Buffer: On-policy learning with parallel exploration
  • Actor-Critic: Learns both policy and value function
  • Diverse Exploration: Each worker explores differently

Famous for: DeepMind’s breakthrough (2016), enables CPU-only training

How It Works

A3C runs multiple agents in parallel, each exploring different strategies. Workers periodically synchronize with a global network, enabling diverse exploration without replay buffers.

Example

// Create an asynchronous A3C agent with parallel workers
var options = new A3COptions<double> { StateSize = 4, ActionSize = 2 };
var agent = new A3CAgent<double>(options);
            
// Select an action given the current environment state
var state = new Vector<double>(new double[] { 0.5, -0.3, 1.0, 0.2 });
var action = agent.SelectAction(state);

Constructors

Constructor Summary
A3CAgent() Initializes a new instance with default settings.
A3CAgent(A3COptions<T> options) Initializes an A3C agent from options.

Properties

Property Summary
int FeatureCount { get; }

Methods

Method Summary
Dictionary<string, T> GetMetrics()
ModelMetadata<T> GetModelMetadata()
T Train() One A3C update step (Mnih et al.
Task TrainAsync() Async wrapper around the sync Train step.
Task TrainAsync(IEnvironment<T> environment, int maxSteps) Train A3C with parallel workers (simplified for single-threaded execution).
Task<Vector<T>> PredictAsync(Vector<T> input)
Vector<T> ComputeGradients(Vector<T> input, Vector<T> target, ILossFunction<T> lossFunction)
Vector<T> Predict(Vector<T> input) IFullModel.Predict surfaces the raw policy-network output (softmax over discrete actions, or [mean | logStd] for continuous actions) rather than the one-hot committed action.
Vector<T> SelectAction(Vector<T> state, bool training)
byte[] Serialize() Auto-generated common model serialization surface.
void ApplyGradients(Vector<T> gradients, T learningRate)
void Deserialize(byte[] data) Auto-generated common model deserialization surface.
void Dispose() Disposes the global networks and every worker’s cached local networks.
void LoadModel(string filepath)
void ResetEpisode()
void SaveModel(string filepath)
void StoreExperience(Vector<T> state, Vector<T> action, T reward, Vector<T> nextState, bool done) Appends one step of (s, a, r, s’, done) experience to the on-policy trajectory buffer.