Dueling Deep Neural Network for Accurate Q Value Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing reinforcement learning systems face challenges in generating reliable advantage estimates, which are computationally complex and require accounting for both environmental states and action-specific properties, leading to difficulties in producing accurate Q values, especially when target values for different actions are close.

Innovation Solution

A dueling deep neural network is employed, featuring separate subnetworks for value and advantage estimates, with a combining layer to generate Q values, allowing for generalized training across actions without altering the underlying reinforcement learning algorithm, and incorporating initial neural network layers like convolutional neural networks for processing observations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate subnetworks are used for value and advantage estimates, then Q value accuracy is improved, but device complexity increases

Engineering Contradiction:
ImproveQ value accuracyVSAvoidnetwork structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The neural network is segmented into separate value subnetwork and advantage subnetwork, each handling different aspects of Q value estimation. The value subnetwork estimates the overall value of a state, while the advantage subnetwork estimates the advantage of specific actions, allowing for more precise Q value calculation through specialized processing paths.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The architecture introduces an additional dimensional separation by processing value and advantage estimates through distinct computational pathways before combining them. This dimensional separation in the network architecture enables independent optimization of each estimation component, improving overall accuracy despite increased structural complexity.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Reliability

If advantage estimates account for both environmental states and action-specific properties, then estimation reliability is improved, but computational complexity increases

Engineering Contradiction:
Improveadvantage estimate reliabilityVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The computational task is segmented into two separate estimation processes: value estimation for environmental states and advantage estimation for action-specific properties. This segmentation allows each subnetwork to focus on specific aspects, improving reliability of individual estimates while managing computational complexity through specialized processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The advantage subnetwork is designed to handle multiple actions simultaneously, providing universal advantage estimates that apply across different actions in a given state. This multi-functionality approach improves reliability by consistent estimation methodology while optimizing computational efficiency through batch processing of multiple actions.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3360083B1Dueling deep neural networks
Publication Date: 2023.10.25 DEEPMIND TECH LTD
  • EP3360083B1 patent drawingFigure 1
  • EP3360083B1 patent drawingFigure 2
  • EP3360083B1 patent drawingFigure 3

AI summary

Systems, methods, and apparatus, including computer programs encoded on a computer storage medium, for selecting an actions from a set of actions to be performed by an agent interacting with an environment. In one aspect, the system includes a dueling deep neural network. The dueling deep neural network includes a value subnetwork, an advantage subnetwork, and a combining layer. The value subnetwork processes a representation of an observation to generate a value estimate. The advantage subnetwork processes the representation of the observation to generate an advantage estimate for each action in the set of actions. The combining layer combines the value estimate and the respective advantage estimate for each action to generate a respective Q value for the action. The system selects an action to be performed by the agent in response to the observation using the respective Q values for the actions in the set of actions.