Automatic Inventory Control: A Neural Network Approach Nicholas Hall ECE 539 12/18/20032TABLE OF CONTENTS INTRODUCTION .......................................................................................3 CHALLENGES ...........................................................................................4 APPROACH................................................................................................6 EXAMPLES ...............................................................................................11 EXPERIMENTS ........................................................................................ 13 RESULTS................................................................................................. 15 CONCLUSION.......................................................................................... 17 REFERENCES.......................................................................................... 183INTRODUCTION Managing an inventory is one of the biggest problems that many businesses face. Whether they manufacture goods, distribute, or resell them, any business that has a physical inventory eventually runs into these problems: • How many of each part should be ordered • Determining when each part should be ordered • How to estimate when demand will go up for a part • How to estimate when demand will go down for a part • How to predict when the entire business’ sales will go up or down Many of these factors are related, but in essence they can be summarized with the rule: Parts should arrive just before a customer orders them, but no sooner. Obviously, this is a very idealized goal, and almost impossible to realize, but many businesses, such as Walmart and Amazon.com, spend millions of dollars each year to try to achieve as close as possible to this. In reality, it is better to accept that it is impossible to predict customer demand 100% and therefore it will be impossible to meet this goal. Backorders can be prevented by filling a warehouse with as many parts as you predict will be sold in 10 years, but then the cost of the inventory would be incredibly high. Inventory cost can be kept down by moving to a Just-In-Time inventory solution, where products are ordered or manufactured immediately after a customer orders them, but this delays every order and leads to dissatisfaction among customers. Therefore, some compromise must be made between the total number of backorders allowed and the average time a particular product is kept in-stock. If this average time a product is kept in-stock is measured in months, then 12 divided by this time measures how many times the warehouse sells all of its products in a year, on average, and is called the “turnover time”.4CHALLENGES BACKGROUND INFORMATION: This project is being done using inventory data from Manufacturer’s Supply. Manufacturer’s Supply is a mail order and Internet business that sells snowmobile, lawnmower, go-kart, and other small engine parts. With a large warehouse and over 30,000 products available for customers to order, it has been a challenge in the past to keep enough products in stock in order to minimize backorders, while at the same time minimizing the average time a product is kept in stock at the warehouse. In the past, inventory management was done almost entirely by a few people who worked on the shipping floor, knew inventory levels and knew how products were moving. As the volume of orders grew, it became increasingly difficult for these people to be aware of current inventory levels and what everyone was ordering. Tools were created in order to help these people quickly visualize current inventory levels and order history for particular parts, which greatly helped them keep up with the trends and understand the inventory. Unfortunately this system still depends on a large human component, and if someone new comes in it would take a long time before they could understand the inventory at the same level. Also, as the volume of sales continues to increase, it will be difficult to fully understand the inventory even with advanced visualization tools. Ideally, a computer program could be developed to replace these human experts and automate the entire parts ordering process. This program would save people time, and would probably increase the accuracy as sales volume increases. PROBLEMS: • Manufacturer’s Supply’s product lines are high seasonal. Several hundreds of certain parts will be sold each month during the winter season, but none will be sold during the summer. Likewise, the weather influences product sales a great deal. In a winter with heavy snowfall they might sell several times the number of snowmobile products than in a winter with little snow. • Only a few of a certain product may sell over a period of several years, making stocking of that product impractical. If in one month someone buys many of these parts unexpectedly, the person ordering parts will realize that their order was one-time and not stock up on these parts that rarely sell. However, it is difficult for a computer program to understand this so the feature set must be designed to differentiate between this unusual, non-recurring “spike”, and the normal selling pattern for this product.5• Product lead times, the time it takes for a product to arrive after it is ordered from a vendor, are not very well defined. Some vendors ship parts so they arrive the very next day, but some vendors require several months to manufacture parts after they are ordered. Likewise, some vendors will sometimes respond with parts in only a matter of days after they are ordered but at other times require over a month to fulfill the order. • Although several years of inventory sales data are available for every product, it is only broken down in sales per month.6APPROACH To predict how to order each product in the inventory, a program and a library were developed to use the multilayer perceptron algorithm (MLP). Because of the large amount of data, the amount of time required to analyze this data, and the likelihood of needing additional features such as the ability to save a network’s state and restart where it left off, it was decided not to use a Matlab approach. Instead, a MLP library was developed from scratch that allows for all the features that this program needs. This library was designed to be generic and not just for this specific problem, and is fully
View Full Document
Unlocking...