Planning Neural Network Implicit Plan Generation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine learning models lack efficient mechanisms for an agent to generate implicit plans that enable goal-directed behavior and exploration in dynamic environments, limiting their ability to generalize and perform tasks effectively in new situations.

Innovation Solution

The system employs a planning neural network that processes planning embeddings using self-attention mechanisms and residual neural network blocks to generate implicit plans, allowing the agent to iteratively update and share information among past interactions, enabling more effective planning and action selection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If traditional machine learning models are used for action selection, then the system is simpler to implement, but the agent cannot generate implicit plans for goal-directed behavior

Engineering Contradiction:
Improvegoal-directed behavior capabilityVSAvoidmodel architecture complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The model is segmented into distinct functional components: an embedding generator that creates planning embeddings from past interactions, a planning neural network with self-attention layers that processes these embeddings to generate implicit plans, and an action selection network that converts plans into actions. This segmentation allows each component to specialize in a specific aspect of planning while maintaining overall system manageability

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimensional representation by creating planning embeddings that capture temporal patterns from past interactions across multiple time steps. This embedding dimension transforms raw interaction data into a compressed representation that the planning network can process efficiently, enabling goal-directed behavior without proportionally increasing complexity

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Productivity

If the agent explores the environment extensively to learn tasks, then it gains better understanding of the environment, but it takes more time steps to accomplish tasks

Engineering Contradiction:
Improvetask completion speedVSAvoidenvironmental understanding
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The planning embeddings are constructed from past interactions and stored in an external memory, preparing the agent with pre-processed environmental understanding before actual task execution. This preliminary organization of knowledge allows the agent to quickly retrieve relevant information during action selection without needing to re-explore the environment

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The planning embeddings serve as an intermediary representation between raw environmental interactions and action selection. These embeddings capture essential environmental patterns and constraints, allowing the agent to make informed decisions without directly experiencing every environmental detail, thus reducing exploration time while maintaining understanding

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If the agent uses detailed representations of current observations in planning embeddings, then it has more information for decision making, but it increases computational complexity

Engineering Contradiction:
Improvedecision making accuracyVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent extracts only the most relevant features from current observations by incorporating them selectively into the planning embeddings through the embedding generator. Rather than processing complete observation representations, the system extracts key environmental cues and integrates them with historical interaction patterns, reducing computational load while maintaining decision accuracy

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The planning embeddings dynamically adjust their representation parameters based on the task at hand and the current observation. The embedding generator modifies which features are emphasized in the planning embeddings depending on the goal, allowing the system to maintain high decision accuracy while adapting computational resources to only process relevant information

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4085385B1Generating implicit plans for accomplishing goals in an environment using attention operations over planning embeddings
Publication Date: 2024.09.11 DEEPMIND TECH LTD
  • EP4085385B1 patent drawingFigure 1
  • EP4085385B1 patent drawingFigure 2
  • EP4085385B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting actions to be performed by an agent interacting with an environment to accomplish a goal. In one aspect, a method comprises: generating a respective planning embedding corresponding to each of multiple experience tuples in an external memory, wherein each experience tuple characterizes interaction of the agent with the environment at a respective previous time step; processing the planning embeddings using a planning neural network to generate an implicit plan for accomplishing the goal; and selecting the action to be performed by the agent at the time step using the implicit plan.