Distributional Q Neural Network for Stable RL Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional reinforcement learning systems select actions using single expected Q returns, which limits their ability to learn a rich action selection model and results in unstable parameter updates, requiring more computational resources and longer training times.

Innovation Solution

The system employs a distributional Q neural network that generates probability distributions over Q returns for action-observation pairs, allowing for more effective action selection and stable parameter updates by training on full Q return probability distributions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If conventional Q neural networks generate single expected Q returns for action-observation pairs, then the system can select actions, but the system cannot learn a rich action selection model and requires more computational resources for training

Engineering Contradiction:
Improveaction selection model richnessVSAvoidcomputational resource consumption
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent changes the parameter representation from single expected Q return values to full probability distribution parameters. The distributional Q network outputs parameters that define a probability distribution over Q returns, capturing uncertainty and variability in action outcomes. This parameter transformation enables richer action selection models while maintaining computational efficiency through targeted distribution parameterization.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If conventional systems train on single expected Q returns, then training can proceed, but parameter updates are unstable and require more training iterations

Engineering Contradiction:
Improveparameter update stabilityVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent implements feedback through the full probability distribution of Q returns, which provides richer information about outcome uncertainty and variability. This feedback mechanism enables more stable parameter updates by capturing the complete distributional characteristics of action outcomes rather than relying on single expected values, leading to faster convergence and reduced training time.

Inventive Principle:
Principle #23Feedback

3Adaptability or versatility

If distributional Q networks model full probability distributions over Q returns, then the system can learn richer action selection models, but the network architecture becomes more complex

Engineering Contradiction:
Improveaction selection effectivenessVSAvoidnetwork architecture complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the Q return space into discrete bins or intervals. The distributional Q network outputs parameters for each bin, representing the probability distribution across different Q return ranges. This segmentation approach enables effective modeling of action selection while managing network complexity through structured discretization of the continuous Q return space.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3933713B1Distributional reinforcement learning
Publication Date: 2025.06.04 GDM HOLDING LLC
  • EP3933713B1 patent drawingFigure 1
  • EP3933713B1 patent drawingFigure 2
  • EP3933713B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting an action to be performed by a reinforcement learning agent interacting with an environment. A current observation characterizing a current state of the environment is received. For each action in a set of multiple actions that can be performed by the agent to interact with the environment, a probability distribution is determined over possible Q returns for the action - current observation pair. For each action, a measure of central tendency of the possible Q returns with respect to the probability distributions for the action - current observation pair is determined. An action to be performed by the agent in response to the current observation is selected using the measures of central tendency.