Multi-Sigma Help

In Multi-Sigma, text data analysis is not supported. If your data contains text data other than in the header section, it may result in a system error. Please make sure your data doesn’t contain any text data when uploading it.

For the trial version, we please request that only one representative from your department registers. Each account can only submit one job at a time, so please coordinate the usage among multiple users to avoid conflicts.

Since the contribution values sum up to 100%, it is possible for a highly contributing parameter to overshadow others. To address this, you can narrow the range of values for highly contributing parameters during analysis, which may make it easier to observe the effects of other parameters.

Multi-Sigma offers a feature to generate experimental conditions randomly. You can conduct experiments using these conditions and combine them with your past experimental data as input. This way, you can effectively use your existing data for AI learning.

While Multi-Sigma has a graphing feature for AI prediction results, it displays only the last few dozen data points. To view the entire dataset, it’s recommended to download the result file and visualize it using tools like Excel.

You can create a two-dimensional graph with the most important objective variable plotted on the x-axis and the next most important one on the y-axis. Alternatively, you can download the multi-objective optimization result file, normalize each objective variable by dividing them by their maximum values, and then visualize the normalized data.

A good starting point is a 9:1 ratio for training and validation data. You can use Excel’s RAND() function for randomization when splitting the data.

If the auto-tuning feature for AI learning doesn’t improve prediction accuracy, consider adjusting how you create the training data, increasing the amount of training data, or understanding that some problems may inherently have low prediction accuracy regardless of the method used.

You can achieve this by including either A or B in the training data. Alternatively, include both A and B in the training data and use constraints during optimization, setting the coefficients of A and B to 1 while ensuring they sum up to 1 (or 100%).

The maximum data size you can upload for one analysis (project) is 3MB. It’s possible that your data may exceed this size. If you encounter this issue, as you mentioned, you can consider reducing the data by removing some portions. We are exploring options for handling larger datasets in future developments.

In Multi-Sigma, you can obtain sets of Pareto solutions for multiple objective variables. You can download these results and use software like Excel to extract the solutions that offer the best balance of objective variables.

Yes, it’s possible. You can convert qualitative data into numerical values. For example, you can assign numerical values like “0” for “Male” and “1” for “Female” to handle gender data.

Example 1:

Age Height Weight Gender

25 172.2 62.7 Male

32 159.3 57.2 Female

48 168.5 65.1 Male

 ↓ Convert Gender to Numerical Values

Age Height Weight Gender

25 172.2 62.7 0

32 159.3 57.2 1

48 168.5 65.1 0

Example 2:

Age Height Weight Blood Type

25 172.2 62.7 A

32 159.3 57.2 B

48 168.5 65.1 O

 ↓ Convert Blood Type to Numerical Values

Age Height Weight A B O

25 172.2 62.7 1 0 0

32 159.3 57.2 0 1 0

48 168.5 65.1 0 0 1

Currently, this functionality is not available. While you can download the AI model you’ve created, it’s intended to be used outside of Multi-Sigma, such as in Python. We are considering the possibility of adding project and AI model backup features in the future.

It’s advisable to include all material candidates in experiments. You can use factor analysis to determine the contributions of each material. During optimization, you can set the maximum and minimum values of materials you don’t want to use to be the same, effectively optimizing only the materials you want to use.

Currently, there’s no built-in feature for this. To compare prediction accuracy, download the prediction results from each model and compare them manually to the ground truth data using tools like Excel. We may consider adding this feature in future developments.

You can use the “Data Profiling” feature to visualize the distribution of your data. This can help you identify trends in missing data and decide what kind of experimental data to collect next to address any biases. Additionally, using Bayesian optimization can prioritize gathering data from sparse regions to improve the overall dataset.

Data with missing values will result in an error. To handle missing values, you should either impute them using an appropriate method or remove rows with missing data before uploading the dataset.

You can analyze data with singular changes, but it’s important to consider the presence of these singularities. Singularities or abrupt changes in the data can significantly impact the analysis and modeling process. It’s crucial to preprocess the data and choose appropriate modeling techniques to handle such cases effectively.

In general, having variables with different scales can negatively impact predictive accuracy in neural networks. Therefore, Multi-Sigma performs data preprocessing to standardize or normalize the numerical scales before analysis.

There are no restrictions. You can use either integers or real numbers (decimal values). When performing optimization, you can choose whether each explanatory variable should be treated as an integer or real number.

In the field of physics, an error of around 10% is often considered high accuracy. In the field of biology, a prediction error of around 20% is considered reasonable.

It’s possible that the page hasn’t been refreshed. Please try refreshing the page by pressing the CTRL key and F5 key simultaneously.

Multi-Sigma uses the Partial Derivative (PaD) method described in Gevrey et al. (2003) to calculate the contribution scores.

Multi-Sigma has been used in the past for optimizing features and specifications of new technologies based on consumer purchase probability from survey data. Currently, it is being used for demand and sales forecasting for aluminum can beverages as part of a NEDO project. However, these cases may not have publicly available materials.

While the job status is “CREATED,” the analysis library is being loaded, and you’ll need to wait. Once the training starts and the job status changes to “RUNNING,” you should be able to click on “Force Terminate” to halt the training.