Off-Policy Actor-Critic Training With Distributed Experience Replay

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing reinforcement learning systems struggle with efficient parallelization and data throughput in training action selection networks, particularly when using off-policy actor-critic techniques, as they often require sequential processing of experience tuples and lack effective distributed training methods.

Innovation Solution

A distributed training system utilizing multiple actor computing units to generate experience tuples, which are processed by learner computing units to train an action selection network using off-policy actor-critic reinforcement learning, enabling parallel processing and efficient training of neural networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If on-policy experience tuples are used for training, then training stability is improved, but sample efficiency deteriorates and training time increases

Engineering Contradiction:
Improvetraining stabilityVSAvoidsample efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent introduces an intermediary mechanism (importance sampling weights and GAE advantages) that mediates between on-policy and off-policy data. This allows the system to use off-policy experience tuples while maintaining training stability through careful weighting and advantage estimation, thereby improving sample efficiency without sacrificing reliability

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If on-policy experience tuples are used for training, then training stability is improved, but computational resources and training time increase

Engineering Contradiction:
Improvetraining stabilityVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-computing importance sampling ratios and using Generalized Advantage Estimation (GAE) to pre-process advantage values from off-policy trajectories. This preliminary processing enables stable training without requiring extensive online interaction, significantly reducing training time while maintaining stability

Inventive Principle:
Principle #10Preliminary action

3Productivity

If off-policy action selection policy is used, then sample efficiency is improved, but training stability deteriorates

Engineering Contradiction:
Improvesample efficiencyVSAvoidtraining stability
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent changes key parameters including using importance sampling ratios to re-weight off-policy gradients, applying GAE to transform raw returns into stable advantage estimates, and using clip ranges to bound policy updates. These parameter changes enable the system to use efficient off-policy samples while maintaining training stability through controlled gradient updates

Inventive Principle:
Principle #35Parameter changes

4Productivity

If off-policy action selection policy is used, then sample efficiency is improved, but computational complexity increases

Engineering Contradiction:
Improvesample efficiencyVSAvoidcomputational complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent extracts and separates the computationally intensive components (importance sampling ratio calculation, GAE advantage estimation) into distinct preprocessing steps. This extraction allows the main training loop to work with pre-computed, stabilized values, reducing the computational complexity during actual policy optimization while maintaining high sample efficiency

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4492291B1Distributed training using off-policy actor-critic reinforcement learning
Publication Date: 2026.05.13 GDM HOLDING LLC
  • EP4492291B1 patent drawingFigure 1
  • EP4492291B1 patent drawingFigure 2
  • EP4492291B1 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 used to select actions to be performed by an agent interacting with an environment. In particular techniques are described for reinforcement learning using off-policy experience tuple trajectories. The off-policy experience tuple trajectories characterize the interaction of the agent with the environment over a sequence of time steps as the agent performed actions selected in accordance with an off-policy action selection policy i.e. one that is different to a current action selection policy of the agent.