Predictron Neural Networks for Adaptive Planning and Value Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional systems separately learn environment models and planning models, resulting in a mismatch that affects the accuracy of value function estimation for predicting outcomes in environments, and require extensive labeled training data.

Innovation Solution

A predictron system that integrates a state representation neural network, a prediction neural network, and a value prediction neural network to jointly learn an environment model and a planning model, allowing for unsupervised training and reducing the need for labeled data, with the ability to adapt the number of planning steps based on internal state representations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If conventional systems separately learn environment models and planning models, then the learning process is simpler and more modular, but the accuracy of value function estimation deteriorates due to model mismatch

Engineering Contradiction:
Improvemodel learning structureVSAvoidvalue function estimation accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent merges the environment model learning and planning model learning into a single integrated neural network system. The network simultaneously learns to predict environment transitions and compute value functions, with shared hidden layers that enable the models to adapt to each other, eliminating the model mismatch problem inherent in separate learning approaches

Inventive Principle:
Principle #5Merging (Combining)

2Ease of manufacture

If conventional systems use separate learning for environment and planning models, then training can proceed independently for each model, but the requirement for labeled training data increases

Engineering Contradiction:
Improvetraining process independenceVSAvoidlabeled training data
Core Design Contradiction:
Ease of manufactureVSQuantity of substance

Solution Approach 1:

The integrated neural network processes training data through shared hidden layers that serve both environment modeling and value function estimation. This merging allows the system to learn from the same training data for both functions simultaneously, reducing the total labeled data required compared to separate learning approaches that would need independent training datasets

Inventive Principle:
Principle #5Merging (Combining)

3Device complexity

If the system uses a fixed number of planning steps, then the computational process is simpler to manage, but the accuracy of outcome prediction deteriorates when the fixed number is not optimal

Engineering Contradiction:
Improveplanning process controlVSAvoidoutcome prediction accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent implements dynamic planning by allowing the neural network to adaptively determine the number of planning steps based on the input state and learned patterns. The system can vary the planning horizon dynamically, using more steps when complex reasoning is needed and fewer steps when the outcome is more directly predictable, thereby optimizing accuracy across different scenarios

Inventive Principle:
Principle #15Dynamics

4Measurement precision

If the system adapts the number of planning steps based on internal state representations, then the accuracy of value function estimation improves, but the computational resources required increase

Engineering Contradiction:
Improvevalue function estimation accuracyVSAvoidcomputational resources
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The dynamic adaptation of planning steps is implemented through the neural network's learned internal state representations, which encode information about the complexity and uncertainty of the prediction task. The system dynamically adjusts computational effort based on these representations, investing more resources only when the internal state indicates that additional planning steps will yield meaningful improvements in accuracy

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3523760B1Reinforcement learning systems
Publication Date: 2024.01.24 DEEPMIND TECH LTD
  • EP3523760B1 patent drawingFigure 1
  • EP3523760B1 patent drawingFigure 2
  • EP3523760B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for prediction of an outcome related to an environment. In one aspect, a system comprises a state representation neural network that is configured to: receive an observation characterizing a state of an environment being interacted with by an agent and process the observation to generate an internal state representation of the environment state; a prediction neural network that is configured to receive a current internal state representation of a current environment state and process the current internal state representation to generate a predicted subsequent state representation of a subsequent state of the environment and a predicted reward for the subsequent state; and a value prediction neural network that is configured to receive a current internal state representation of a current environment state and process the current internal state representation to generate a value prediction.