Recurrent Environment Predictors Using Neural Network State Extraction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional neural network systems require generating predicted observations at every time step to produce accurate future predictions, which is inefficient and consumes more computational resources, resulting in less accurate, spatially, and temporally incoherent predictions.

Innovation Solution

A neural network system that includes a recurrent neural network updating its hidden and cell states based on preceding actions and a decoder neural network processing these states to generate predicted observations at future time steps, allowing for efficient and accurate high-dimensional predictions without generating intermediate observations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If predicted observations are generated at every time step, then accurate future predictions are achieved, but computational resource usage increases and efficiency decreases

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent extracts only the essential information needed for future predictions by using a recurrent neural network to maintain hidden states that summarize past observations, rather than processing and storing every intermediate prediction. This allows the system to generate accurate long-term predictions without computing every intermediate time step, thereby improving computational efficiency while maintaining prediction accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

2Measurement precision

If predicted observations are generated at every time step, then accurate future predictions are achieved, but computational resource usage increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational resource usage
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system extracts only the necessary predictive information by using recurrent neural network hidden states to compress and represent the essential features of past observations. This extraction approach enables accurate future predictions without the computational overhead of generating and processing predictions at every intermediate time step, thereby reducing computational resource usage while maintaining accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

3Productivity

If conventional neural network systems generate predictions at every time step, then future predictions are produced, but spatial and temporal coherence deteriorates

Engineering Contradiction:
Improveprediction generation capabilityVSAvoidspatial and temporal coherence
Core Design Contradiction:
ProductivityVSStability of the object's composition

Solution Approach 1:

The patent applies preliminary action by using the recurrent neural network to pre-process and encode sequential observations into compressed hidden states before generating predictions. This preliminary encoding maintains the temporal relationships and spatial coherence of the environment state, allowing the system to generate coherent predictions for future time steps without the degradation that occurs when processing every intermediate step.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3523761B1Recurrent environment predictors
Publication Date: 2023.09.20 DEEPMIND TECH LTD
  • EP3523761B1 patent drawingFigure 1
  • EP3523761B1 patent drawingFigure 2
  • EP3523761B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for environment simulation. In one aspect, a system comprises a recurrent neural network configured to, at each of a plurality of time steps, receive a preceding action for a preceding time step, update a preceding initial hidden state of the recurrent neural network from the preceding time step using the preceding action, update a preceding cell state of the recurrent neural network from the preceding time step using at least the initial hidden state for the time step, and determine a final hidden state for the time step using the cell state for the time step. The system further comprises a decoder neural network configured to receive the final hidden state for the time step and process the final hidden state to generate a predicted observation characterizing a predicted state of the environment at the time step.