Paper notes: Dueling Network architectures for deep reinforcement learning

Dueling Network architectures for deep reinforcement learning

ICML Best Paper

Google DeepMind

• Abstract:

This article is one of ICML 2016 's best papers and is also from Google DeepMind.

In recent years, on the reinforcement learning on the deep representation have achieved great success. However, many of these applications take advantage of traditional network architectures, such as neural networks, LSTMS, or auto-encoders. In this paper, a new network structure is proposed to deal with model-free reinforcement learning. The proposed dueling network represents two independent predictions:

One for the state value function;

One for the State-dependent action Advantage function.

The main advantages of this decomposition approach are: generalize learning across actions without imposing any change to the underlying reinforcement learning Algo Rithm. (A learning crossover is generated without any change to the potential RL). Experiments show that when there are many similar values, we can learn a better strategy evaluation. In addition, when playing Atari 2600 made a better effect than the Nature.

　　

　　

Paper notes: Dueling Network architectures for deep reinforcement learning