Multi-Sigma: Simultaneously predicting multiple outputs and performing multi-objective optimization through inverse analysis.
The recent advancements in AI tools have been remarkable, leading to the emergence of numerous no-code applications for leveraging machine learning and deep learning, commonly referred to as Auto ML. Currently, most machine learning tools available are designed for predictive purposes and typically analyze only one output. However, considering various real-world application scenarios, this single-output approach is often insufficient.
For instance, in sales and marketing, multiple outputs such as customer satisfaction metrics, sales figures, and net profit are often needed. Similarly, in manufacturing, multiple evaluation criteria like quality, cost, and functionality are common. Analyzing data with multiple outputs using existing Auto ML tools requires preparing separate training data and conducting analysis for each output.
Furthermore, in optimization scenarios, such as maximizing customer satisfaction through sales activities or product specifications, or optimizing manufacturing conditions to simultaneously improve quality, cost, and multiple functionalities, there is a clear need for tools that can handle multiple outputs. After making predictions, the next logical step is to explore optimal conditions, which is a general requirement.
Multi-Sigma is an extremely unique tool designed to analyze multiple outputs simultaneously and perform inverse analysis on them. With Multi-Sigma, up to 200 input parameters and 100 output parameters can be analyzed, allowing for simultaneous prediction of 100 outputs from 200 inputs and the exploration of optimal conditions for the 200 inputs based on the 100 outputs. Additionally, multiple constraints can be added to the inputs, a feature not found in any other tool worldwide. This article will explain why other tools find it challenging to achieve this.
Many analytical methods cannot handle multiple outputs.
First, most of the analytical methods available today can only handle a single output. In contrast, neural networks, a method of deep learning, are capable of handling multiple outputs, making them a highly unique approach. Therefore, tools that can manage multiple outputs are primarily based on neural network technology.
Prediction and optimization are entirely different algorithms.
It is important to understand that prediction and optimization generally involve different algorithms. Therefore, performing both prediction and optimization requires the integration of two distinct algorithms. Recently, Bayesian optimization has emerged as a highly useful technique, rapidly gaining popularity, especially in fields like materials informatics.
At first glance, Bayesian optimization seems to perform optimization directly. However, it often involves creating a predictive model using Gaussian regression, which is a method within deep learning, followed by optimization using quasi-Newton methods. Since Gaussian regression fundamentally can handle only one output, current tools capable of optimizing a single output often rely on Bayesian optimization techniques. Although recent methods have been proposed to integrate multiple outputs into a single metric for analysis via Bayesian optimization, this topic will be covered in a separate discussion.
Predictive AI × Optimization AI
While discussing neural networks and Bayesian optimization (Gaussian regression), it’s worth noting that these two deep learning techniques represent the two major branches of the field. In my opinion, mathematicians and theorists tend to favor Bayesian optimization, and it’s curious that few people use both. Personally, I feel that the strengths and weaknesses of both methods complement each other, making it essential to use them appropriately depending on the application (this will be elaborated in another article).
Multi-Sigma allows for switching between neural network analysis and Gaussian regression, the two major predictive methods in deep learning, for analysis. It also integrates these predictive methods with genetic algorithms, a well-known optimization technique. This integration enables simultaneous prediction of multiple outputs and inverse analysis for multiple outputs. Since genetic algorithms are also considered a part of AI, this approach allows the optimization AI to use predictive AI to automatically search for optimal conditions.
By leveraging both methods, Multi-Sigma can compensate for their respective shortcomings, resulting in a versatile tool applicable to various real-world scenarios.
