In recent years there have been many successes of using deep representations
in reinforcement learning. Still, many of these applications use conventional
architectures, such as convolutional networks, LSTMs, or auto-encoders. In this
paper, we present a new neural network architecture for model-free
reinforcement learning. Our dueling network represents two separate estimators:
one for the state value function and one for the state-dependent action
advantage function. The main benefit of this factoring is to generalize
learning across actions without imposing any change to the underlying
reinforcement learning algorithm. Our results show that this architecture leads
to better policy evaluation in the presence of many similar-valued actions.
Moreover, the dueling architecture enables our RL agent to outperform the
state-of-the-art on the Atari 2600 domain.
[1511.06581] Dueling Network Architectures for Deep Reinforcement Learning