Automatic Inventory Control: A Neural Network Approach

1
Automatic Inventory ControlA Neural Network
Approach

• Nicholas Hall
• ECE 539 Final Project
• Fall 2003

2
Managing Inventory

• Managing inventory is a huge problem for many
  businesses
• How many parts do you order?
• When do you order?
• How do you estimate demand?
• Parts should arrive just before a customer orders
  them, but no sooner!

3
Managing Inventory

• It is impossible to predict customer demand 100
  in almost every case.
• Instead try to
• Minimize backorders - orders for parts that are
  not in stock
• Maximize inventory turnover - number of times a
  year, on average, every product in the warehouse
  is sold.
• Some tradeoff must be made.

4
Prediction Problems

• Many products are highly seasonal

5
Prediction Problems

• Some have impossible to predict buying spikes

6
Prediction

• Fortunately, others show similar patterns each
  year

7
Approach

• Use inputs that reflect the data
• sold for past 12 months
• sold for past 12 months / sold for 12 months
  prior
• sold for seasons / sold for other seasons
• sold for 1 month / sold all year

8
Approach

• Accuracy can never be 100 on the testing set,
  but that is not what we are trying to accomplish.
• Instead, try to minimize backorders while
  maximizing the turnover rate.
• Created simulation program that pretends it is
  managing the inventory for an entire year and
  keeps statistics about product movements.

9
MLP Results

• Using the multi-layer perceptron algorithm with
  150 hidden nodes, best compromise was a 54.9
  backorder rate with 6.7 turns per year.
• Common backorder rate when inventory is managed
  by humans is 2-5, so the MLP did not perform
  very well.

10
Simulated Annealing Results

• Originally, a simulated annealing prediction
  program was developed as a reference.
• It was expected to be outperformed by the MLP.
• However, it achieved a very good 9.3 backorder
  rate with 8.5 inventory turns per year.

11
Conclusion

• The more complex method is not always better, as
  the much simpler simulated annealing program was
  better at predicting that the MLP.
• However, both are slow. The simulated annealing
  program used 1 week of CPU time on an AMD Athlon
  1800.
• If the amount of training time for the MLP is
  increased, it may do better.