RandomNetworkDistillation<T>
Models & Types · AiDotNet.ReinforcementLearning.IntrinsicMotivation
Random Network Distillation (Burda et al. 2018): intrinsic reward is the error of a trained predictor network against a fixed, randomly-initialized target network. Novel states have high prediction error (the predictor has not learned them yet); familiar states have low error, so the bonus fades.
Implements: IIntrinsicRewardModule<T>
How It Works
Both networks are single-hidden-layer MLPs (tanh). The target is fixed random; the predictor is trained by one SGD step per visited state to match the target’s output on that state. The intrinsic reward is the mean-squared prediction error, normalized by a running estimate so its scale stays stable as training proceeds. Networks are lazily sized to the first state seen.
Constructors
| Constructor | Summary |
|---|---|
RandomNetworkDistillation(int hiddenSize, int outputSize, double learningRate, int? seed) |
Creates an RND intrinsic-reward module. |
Methods
| Method | Summary |
|---|---|
T ComputeIntrinsicReward(Vector<T> state) |
|
void Reset() |
|
void Update(Vector<T> state) |