RainbowDQNAgent<T>
Models & Types · AiDotNet.ReinforcementLearning.Agents.Rainbow
Rainbow DQN agent combining six extensions to DQN.
Inheritance: ReinforcementLearningAgentBase<T> → DeepReinforcementLearningAgentBase<T> → RainbowDQNAgent<T>
Implements: IGeneratedParameterRegistrar<T>, IGradientComputable<T, Vector<T>, Vector<T>>
For Beginners
Rainbow takes the best ideas from six different DQN improvements and combines them. It’s currently the strongest DQN variant, achieving state-of-the-art performance.
Six components:
- Double Q-learning: Reduces overestimation
- Dueling Architecture: Separates value and advantage
- Prioritized Replay: Samples important experiences more
- Multi-step Returns: Better credit assignment
- Distributional RL (C51): Learns distribution of returns
- Noisy Networks: Parameter noise for exploration
Famous for: DeepMind’s combination achieving human-level Atari performance
How It Works
Rainbow combines: Double Q-learning, Dueling networks, Prioritized replay, Multi-step learning, Distributional RL (C51), and Noisy networks.
Example
// Create a Rainbow DQN agent combining six DQN improvements
var options = new RainbowDQNOptions<double> { StateSize = 4, ActionSize = 2, NSteps = 3 };
var agent = new RainbowDQNAgent<double>(options);
// Select an action using distributional RL with noisy exploration
var state = new Vector<double>(new double[] { 0.5, -0.3, 1.0, 0.2 });
var action = agent.SelectAction(state);
Constructors
| Constructor | Summary |
|---|---|
RainbowDQNAgent() |
Initializes a new instance with default settings. |
RainbowDQNAgent(RainbowDQNOptions<T> options) |
Initializes a Rainbow DQN agent from options. |
Properties
| Property | Summary |
|---|---|
int FeatureCount { get; } |
Methods
| Method | Summary |
|---|---|
Dictionary<string, T> GetMetrics() |
|
ModelMetadata<T> GetModelMetadata() |
|
T Train() |
|
Vector<T> ComputeGradients(Vector<T> input, Vector<T> target, ILossFunction<T> lossFunction) |
Computes gradients of the loss with respect to this agent’s parameters, without updating them. |
Vector<T> Predict(Vector<T> input) |
IFullModel.Predict surfaces the raw Q-value vector (one element per discrete action) 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 LoadModel(string filepath) |
|
void ResetEpisode() |
|
void SaveModel(string filepath) |
|
void StoreExperience(Vector<T> state, Vector<T> action, T reward, Vector<T> nextState, bool done) |
|
void Train(Vector<T> state, Vector<T> target) |
Supervised Train(state, target) does one direct regression step on the online Q-network: minimise ‖Q(state, ·) − target‖² so callers that provide an explicit Q-value target — offline pretraining from logged data, distillation, BC warm start… |