Neural Architecture Search via Hybrid Reinforcement and Random Mutation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural architecture search (NAS) methods face challenges in efficiently exploring the vast search space of neural network model structures, leading to suboptimal performance and model degradation during the evolution process.

Innovation Solution

The proposed method combines reinforcement learning mutation and random mutation to generate new neural network model structures. It starts with an initial population generated based on multi-objective optimization hyperparameters, performs selection and crossover, and then applies reinforcement learning mutation and random mutation to create new structures. This process continues until a multi-objective optimization state is optimal, allowing for the selection of a neural network model structure suitable for different scenarios.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If reinforcement learning mutation is used to generate new neural network structures, then the search efficiency and optimization performance are improved, but the complexity of the search process and computational resources required increase

Engineering Contradiction:
Improvesearch efficiencyVSAvoidsearch process complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent combines reinforcement learning mutation with random mutation in a hybrid approach. The reinforcement learning agent generates mutations based on learned policies while random mutation provides exploration. This merging allows the system to achieve high search efficiency through learned optimizations while managing complexity by using random mutation as a simple baseline for comparison and hybridization.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent introduces a population-based framework where multiple candidate structures are maintained and evaluated. This population acts as an intermediary between the reinforcement learning agent and the final selected structure. The agent operates on the population level, managing complexity by distributing the search across multiple candidates rather than directly optimizing a single structure.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If a large population size is used in the evolutionary algorithm, then the coverage of the search space is improved, but the computational time and resources required increase

Engineering Contradiction:
Improvesearch space coverageVSAvoidcomputational time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent employs dynamic population management where the population size and composition are adjusted during the evolution process. The reinforcement learning agent can focus computational resources on promising regions of the search space while maintaining a manageable population size. The population dynamically adapts to the current state of search, balancing coverage with computational efficiency.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes parameters such as population size, mutation rate, and selection pressure during the evolution process. By dynamically adjusting these parameters, the system can maintain adequate search space coverage while reducing computational time. The reinforcement learning agent guides these parameter changes based on learned optimal strategies.

Inventive Principle:
Principle #35Parameter changes

3Manufacturing precision

If multi-objective optimization is applied to neural architecture search, then the quality of generated models is improved, but the complexity of the optimization process increases

Engineering Contradiction:
Improvemodel qualityVSAvoidoptimization process complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments the multi-objective optimization into separate evaluation criteria and uses the reinforcement learning agent to handle each objective systematically. The agent learns to balance multiple objectives by receiving rewards based on different model qualities. This segmentation allows complex multi-objective optimization to be managed through a modular, learnable process rather than a monolithic complex system.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements feedback mechanisms where the reinforcement learning agent receives rewards based on multi-objective evaluation of generated structures. This feedback loop guides the agent to produce high-quality models that satisfy multiple objectives simultaneously. The feedback system simplifies the optimization process by providing clear directional guidance rather than requiring complex explicit optimization algorithms.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3809333B1Search method, device and storage medium for neural network model structure
Publication Date: 2025.02.26 BEIJING XIAOMI INTELLIGENT TECH CO LTD
  • EP3809333B1 patent drawingFigure 1
  • EP3809333B1 patent drawingFigure 2
  • EP3809333B1 patent drawingFigure 3

AI summary

A search method for a neural network model structure, includes: generating(S11) an initial generation population of network model structure based on multi-objective optimization hyper parameters, as a current generation population of network model structure; performing(S12) selection and crossover on the current generation population of network model structure; generating(S13) a part of network model structure based on reinforcement learning mutation, and generating(S15) a remaining part of network model structure based on random mutation on the selected and crossed network model structure; generating a new population of network model structure based on the part of network model structure generated by reinforcement learning mutation and the remaining part of network model structure generated by random mutation; and searching(S16) a next generation population of network model structure based on the current generation population of network model structure and the new population of network model structure.