Explanation of Logistics AI Model "EIQ_AI": Integration of Shipping Data, Fractal Theory, and SDCA Regression

This page explains the mechanics from data generation to forecasting within our proprietary AI model "EIQ_AI". We introduce a groundbreaking approach to logistics AI that achieves high-precision predictions without relying on massive historical records by combining a single set of shipping data, fractal theory, and SDCA regression.

1. Difference Between Standard AI Models and Logistics AI "EIQ_AI" (Uniqueness of Approach)

Standard machine learning adopts an "inductive" approach, collecting massive past performance data to find patterns.

However, TCalc EIQ_AI takes the unique approach of "using only a single set of shipping data." It utilizes a "deductive" simulation-based method (data augmentation) that generates virtual future scenarios based on theory and rules, then trains the AI model of the logistics AI.

2. Steps for Creating AI Model Data and Application of Fractal Theory

The data for training the AI model in this system is created based on the following steps and theories:

① Application of Fractal Theory
Based on fractal theory, which states that "distribution centers are fractals, and shipping data is also fractal." This principle suggests that even as the scale of logistics data expands or contracts, its internal distribution structure (self-similarity) is preserved.
② Data Augmentation via Fluctuations in E (Destinations) and I (Items)
Using a single baseline shipping data set as a seed, we intentionally vary the destination count (E) and item count (I) to generate approximately 200 data patterns for the AI model.
③ Adding Fluctuations via Monte Carlo Method
When varying E and I, parameters are specified using the "Monte Carlo method" rather than simple uniform multiplication. This introduces probabilistic fluctuation for each item, giving the logistics AI the capability to handle irregular, distorted peaks (such as extreme peak seasons).

【Explanation】Why Can an AI Model Be Built from "a Single Set of Shipping Data"?

Normally, AI requires tens of thousands of past data records, but because this system incorporates "physical laws of logistics operations (domain knowledge)" as calculation rules, scaling (expanding/contracting) a single set of high-quality historical data is sufficient to generate high-precision simulation data (synthetic data) for unknown future volumes.

3. "Physical Laws of Logistics" Proven in the Data Generation Process

Through roughly six months of research on AI model data creation and actual calculation verification, the following lawfulness has been confirmed:

Relationship Between Specified vs. Actual E and I Counts:
When specifying E and I counts to fluctuate during the actual calculation phase of model creation, actual E and I counts on model data decrease according to a specific mathematical law. This occurs due to "duplication" caused by random assignments.
Non-linear Multiplicative Law of Shipping Volume (Q):
Fluctuations in E and I counts act on shipping volume like a "multiplication (area)" operation:
- Destination Count E (0.5x) × Item Count I (0.5x) = Shipping Volume becomes 0.25x (one-fourth).
- Destination Count E (2x) × Item Count I (2x) = Shipping Volume becomes 4x.

4. Bias and Correction in Logistics AI (SDCA Regression) Predictions

When training machine learning (SDCA regression) on the generated model data to run forecast simulations, a phenomenon occurs where **calculated values consistently output approximately 12% to 13% lower than actual calculations**.

【Explanation】Why Does the AI Model (SDCA Regression) Predict "Low"?

Shipping volume (Q) expands along a "multiplicative curve" of E × I, but the adopted SDCA regression algorithm attempts to forecast using an "additive line (linear model)." Because it draws a straight line against a curve, it structurally passes below the curve overall, producing an underestimation bias.

Solution: Introduction of Correction Factor (1.14)

To resolve this discrepancy unique to logistics AI, the system incorporates the following response:

A practical solution is adopted that applies a default multiplier of **1.14 (approx. +12.5% positive correction)** to prediction results.
This correction factor is not fixed, allowing the **user (learner) to modify it arbitrarily**.

Through hands-on operation, learners come to understand the essence of data science and logistics engineering: "rather than blindly trusting values produced by **AI models**, one must understand algorithm tendencies and apply proper human correction."