Manager-Worker Neural Network for Long-Timescale Credit Assignment

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Reinforcement learning systems face challenges in effectively performing long timescale credit assignment, requiring significant computational resources and struggling to recognize the influence of previous states and actions on rewards, especially when using absolute goals rather than directional goals.

Innovation Solution

An action selection system is developed with a decoupled manager-worker structure and a dilated LSTM neural network, which processes intermediate representations to generate latent and action embedding vectors, allowing for long timescale credit assignment and reducing computational requirements by using directional goals instead of absolute goals.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If reinforcement learning systems use traditional action selection methods, then they can select actions based on current state, but they struggle with long timescale credit assignment and require significant computational resources

Engineering Contradiction:
Improvelong timescale credit assignmentVSAvoidcomputational resources
Core Design Contradiction:
Loss of timeVSDevice complexity

Solution Approach 1:

The patent segments the action selection process into two distinct neural network components: a manager network that selects abstract goals and a worker network that selects concrete actions. This segmentation allows the manager to operate on longer timescales for credit assignment while the worker handles immediate action selection, reducing the computational burden on a single system and enabling effective long-timescale learning.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediate representation layer that transforms observations into a compressed latent space before processing. This intermediate representation acts as a mediator between the raw environment state and the neural network processing, reducing computational complexity while preserving essential information for long-timescale credit assignment.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If reinforcement learning systems use absolute goals, then they can provide clear target directions, but they increase computational requirements and reduce training efficiency

Engineering Contradiction:
Improvetraining speedVSAvoidcomputational resources
Core Design Contradiction:
ProductivityVSUse of energy by moving object

Solution Approach 1:

Instead of using absolute goals that specify exact target positions, the patent inverts the approach by using relative directional goals that indicate preferred movement directions. This inversion reduces the computational burden of calculating precise distances and positions while maintaining effective guidance for the agent, thereby improving training efficiency and reducing computational resource consumption.

Inventive Principle:
Principle #13The other way round (Inversion)

3Measurement precision

If reinforcement learning systems process full state representations, then they can make accurate action selections, but they consume more computational resources

Engineering Contradiction:
Improvestate representation accuracyVSAvoidcomputational resources
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent extracts only the essential features from full state representations by projecting them into a compressed latent space. This extraction process removes redundant information while preserving the critical state characteristics needed for accurate action selection, thereby reducing computational resource consumption while maintaining decision-making accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameter representation of the state from high-dimensional raw observations to lower-dimensional latent representations. This parameter transformation maintains the essential information needed for accurate action selection while significantly reducing the computational resources required for processing and storing state information.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3568810B1Action selection for reinforcement learning using neural networks
Publication Date: 2023.08.16 DEEPMIND TECH LTD
  • EP3568810B1 patent drawingFigure 1
  • EP3568810B1 patent drawingFigure 2
  • EP3568810B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for a system configured to select actions to be performed by an agent that interacts with an environment. The system comprises a manager neural network subsystem and a worker neural network subsystem. The manager subsystem is configured to, at each of the multiple time steps, generate a final goal vector for the time step. The worker subsystem is configured to, at each of multiple time steps, use the final goal vector generated by the manager subsystem to generate a respective action score for each action in a predetermined set of actions.