Neural Network Generation via Evolutionary Reinforcement Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network structures are manually designed, requiring significant manual expertise, computing resources, and time, and lack adaptability across different tasks, leading to inefficiencies in generation and poor transportability.
Innovation Solution
A method combining evolutionary and reinforcement learning algorithms to automatically generate and update neural network structures, optimizing and stabilizing the generation process by replacing random variation with feedback-based updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual design methods are used to create neural network structures, then expert knowledge can be applied to optimize performance, but the process requires significant time, computing resources, and lacks adaptability to different tasks
Solution Approach 1:
The system enables neural networks to automatically generate and optimize their own structures through self-play reinforcement learning. The generated networks serve as both players and designers, eliminating the need for manual expert intervention while maintaining high performance through autonomous evolution and adaptation.
Solution Approach 2:
A game-based intermediary environment is introduced where neural networks compete and collaborate to evaluate and refine architectural designs. This game framework acts as a mediator that translates design quality into measurable outcomes, enabling automated selection and optimization without human expertise.
2Reliability
If manual design methods are used to create neural network structures, then specific task optimization is possible, but transportability across different tasks is poor
Solution Approach 1:
The system generates universal neural network architectures that can adapt to multiple tasks through self-play reinforcement learning. The networks learn transferable design patterns and principles that enable them to perform well across different applications, achieving both task-specific optimization and broad versatility simultaneously.
Solution Approach 2:
The neural network designs are made dynamic and adaptive rather than static. The system continuously evolves network architectures through reinforcement learning, allowing the same generation process to adapt to different tasks and requirements, thereby achieving high transportability across diverse applications.
3Adaptability or versatility
If evolutionary algorithms with random variation are used for neural network generation, then exploration of design space is achieved, but stability and speed of generation are insufficient
Solution Approach 1:
The system implements feedback loops where the performance outcomes of generated networks are fed back into the generation process. Reinforcement learning agents use this feedback to learn from successful and unsuccessful designs, progressively improving generation quality and speed while maintaining thorough design space exploration through guided search.
Solution Approach 2:
The patent replaces traditional mechanical evolutionary algorithms with reinforcement learning-based generation. This substitution transitions from random variation driven by mechanical evolution to intelligent, feedback-driven design optimization, significantly improving both the speed and stability of neural network generation while maintaining exploration capability.
Data Source
AI summary
Disclosed are a method for generating a neural network, an apparatus thereof, and an electronic device. The method includes: obtaining an optimal neural network and a worst neural network from a neural network framework by using an evolutionary algorithm; obtaining an optimized neural network from the optimal neural network by using a reinforcement learning algorithm; updating the neural network framework by adding the optimized neural network into the neural network framework and deleting the worst neural network from the neural network framework; and determining an ultimately generated neural network from the updated neural network framework. In this way, a neural network is optimized and updated from a neural network framework by combining the evolutionary algorithm and the reinforcement learning algorithm, thereby automatically generating a neural network structure rapidly and stably.