Distributional Q Network for Continuous Control Stability

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Reinforcement learning systems face challenges in training neural networks to achieve effective control policies for robots in high-dimensional, continuous action spaces, particularly in tasks requiring precise motor control, as they struggle with modeling randomness and uncertainty in continuous action environments.

Innovation Solution

Incorporating a distributional Q network into the training process provides a more stable learning signal by modeling the randomness inherent in interacting with environments, and distributing and parallelizing the training across multiple actor computing units improves training efficiency and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a conventional critic network is used in reinforcement learning training, then the training process is simpler, but the learning signal is less stable and performance is lower

Engineering Contradiction:
Improvelearning signal stabilityVSAvoidnetwork architecture complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The value function is segmented into multiple discrete bins representing different return ranges. Instead of predicting a single continuous value, the critic network predicts a probability distribution over these bins, capturing the uncertainty and randomness in continuous action spaces more effectively.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The output parameters of the critic network are changed from a single continuous Q-value to a probability distribution over discrete bins. This parameter transformation enables better modeling of randomness while maintaining computational tractability through the use of binning and softmax activation.

Inventive Principle:
Principle #35Parameter changes

2Productivity

If training is performed on a single computing unit, then the system is simpler to implement, but the wall clock time for training is longer

Engineering Contradiction:
Improvetraining speedVSAvoidsystem architecture complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The training system is segmented into multiple actor computing units that operate in parallel. Each actor independently interacts with the environment and contributes to the training data, distributing the computational load and reducing wall clock time.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Multiple actors perform preliminary exploration and data collection in parallel before the main training phase. This preliminary action generates diverse training samples that accelerate convergence during subsequent training iterations.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If function approximation is used in continuous action space, then the system can handle high-dimensional actions, but inherent uncertainty and randomness increase

Engineering Contradiction:
Improvecontinuous action space handlingVSAvoidprediction uncertainty
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The continuous value space is segmented into discrete bins, transforming the continuous function approximation problem into a discrete classification problem. This segmentation reduces the uncertainty inherent in continuous function approximation while maintaining the ability to handle high-dimensional action spaces.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The binned probability distribution acts as an intermediary between the continuous action space and the policy network. It mediates the uncertainty by providing a structured, discrete representation that captures statistical properties without the full complexity of continuous function approximation.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3688675B1Distributional reinforcement learning for continuous control tasks
Publication Date: 2024.05.08 DEEPMIND TECH LTD
  • EP3688675B1 patent drawingFigure 1
  • EP3688675B1 patent drawingFigure 2
  • EP3688675B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training an action selection neural network that is used to select actions to be performed by a reinforcement learning agent interacting with an environment. In particular, the actions are selected from a continuous action space and the system trains the action selection neural network jointly with a distribution Q network that is used to update the parameters of the action selection neural network.