Data Processing Apparatus for Boltzmann Machine Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The population annealing method for solving large-scale combinatorial optimization problems faces a high computational cost due to the need for a large number of replicas, leading to inefficiencies in the resampling process and increased computational time.
Innovation Solution
A data processing apparatus that transforms combinatorial optimization problems into Boltzmann Machine evaluation functions, utilizing a population annealing method with a resampling process that duplicates low-energy replicas and eliminates high-energy ones, reducing the need for recalculating local fields and improving computational efficiency by storing and updating values efficiently across replicas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the population annealing method is used to solve large-scale combinatorial optimization problems, then the solution search capability is improved, but the computational cost increases due to the need for a large number of replicas
Solution Approach 1:
The patent creates copies (replicas) of the Boltzmann machine state to enable parallel computation. Multiple replicas are maintained and processed simultaneously, allowing the system to explore multiple solution paths in parallel. This copying approach is fundamental to the population annealing method, where replicas are duplicated, modified through MCMC moves, and then selectively retained or eliminated based on their energy values.
Solution Approach 2:
The patent segments the computational process into distinct phases: MCMC move generation, energy calculation, acceptance/rejection decisions, and resampling. By dividing the overall optimization process into these manageable segments, the system can efficiently manage multiple replicas through structured operations. The resampling process itself segments the population into survivors and eliminated replicas, enabling controlled population management.
2Measurement precision
If many replicas are used in the population annealing method, then the resampling process becomes more thorough, but the computational time increases
Solution Approach 1:
The patent performs preliminary calculations of local fields and energy changes before executing MCMC moves. By pre-computing the necessary information (local fields for each variable), the system avoids redundant calculations during the actual move execution and resampling phases. This preliminary action significantly reduces the computational burden when handling large numbers of replicas.
Solution Approach 2:
The patent implements a resampling mechanism where replicas are temporarily discarded (eliminated) based on their energy values, and high-quality replicas are recovered through duplication. This selective discarding and recovering process maintains population quality while controlling computational resources. The system discards low-energy replicas and recovers/broadcasts high-energy replicas to multiple positions, efficiently managing the replica population.
3Measurement precision
If local fields are recalculated for each replica during resampling, then the accuracy is maintained, but the computational efficiency decreases
Solution Approach 1:
The patent computes local fields in a universal manner that serves multiple purposes: they are used for MCMC move generation, energy calculation, and resampling decisions. By calculating local fields once and reusing them across different computational stages and for multiple replicas, the system maintains accuracy while avoiding redundant calculations. The local fields represent a multi-functional data structure that supports various operations without requiring recalculation.
Solution Approach 2:
The patent merges the computation of local fields across multiple replicas by recognizing that identical configurations share identical local fields. Instead of calculating local fields independently for each replica, the system combines computations by identifying and reusing common field values. This merging approach significantly reduces the total computational effort while maintaining the precision needed for accurate resampling decisions.
Data Source
AI summary
A first storage unit holds the values of discrete variables included in an evaluation function and the values of local fields for each replica. A second storage unit provided for each replica holds the values of corresponding discrete variables and local fields. A processing unit repeats, for each replica, a process of updating the value of any discrete variable, the value of the evaluation function, and the values of the local fields on the basis of a set temperature and the values of the local fields stored in the second storage unit, and after every predetermined number of iterations, performs resampling of population annealing. When a first replica is duplicated to create a second replica, the processing unit reads the values of the discrete variables and local fields of the first replica from the first storage unit and stores them in the second storage unit for the second replica.