In DQN, the online network predicts action values and supports action selection. It learns by matching its predictions for sampled state–action pairs to targets computed with a separate target network, avoiding a target that shifts with every update. section 8.3 section 8.3
Online network
The currently updated neural network in DQN that estimates action values and supports action selection.
How Applied Reinforcement Learning covers it
- How deep it goes
- Explains prediction and regression, while assigning target computation to a delayed network updated periodically.
- Examples
- Illustrates training with sampled transitions, predicted Q-values, Bellman targets, and squared error.
- What you can do after
- Understand how action selection uses online predictions while learning targets come from a delayed network.
- How it connects
- Pairs the online network with Q-learning, replay buffers, Bellman targets, and a target network.
Related concepts
- networkkind of · A network is a set of interconnected nodes and links that represents relationships or enables communication and exchange among the nodes.
- actor networkcontrasts with · An actor network is a learned reinforcement-learning model that represents a policy by producing action choices or action probabilities for given states.
This explainer was generated by Manning's knowledge graph from the text of Applied Reinforcement Learning. The section links point to the parts of the book it draws on.
