During a hot summer day, many people stop at the Rita’s in Nazareth to enjoy a sweet, cool treat. Gallons of ice are eaten but still gallons are not. At the end of the day, gallons upon gallons of perfectly good spare Italian Ice must be thrown away.
Miles away in Easton, Rita’s probably doesn’t have this issue, right?
They still do.
Corporate guidelines force Rita's nationwide to throw away perfectly good ice. Countless Rita’s workers have seen what happens to the spare ice yet turn a blind eye.
Adam Kollgaard is one of these Rita’s workers, but knows something has to be done. Nobody else seemed to care enough to change it, so Kollgaard knew he had to be the one to change it.
Previous Experience
From the spring of his sophomore year to the fall of his junior year, Kollgaard worked at the Rita’s in Nazareth. He had to balance his time working, being at school, and doing extracurriculars such as track. Things got easier for him when school let out and he started to work on his own personal projects.
Simply put, Kollgaard likes to code. He is constantly either looking for something to program or working on a program. His passion started in middle school when he first learned about artificial intelligence and “got really into it.”
At this point he had only completed pre-algebra and couldn't properly understand the complicated calculus required to make artificial intelligence. Instead of spending months learning the algebra and calculus that he wouldn’t learn for years, he simply moved onto other projects.
In high school Kollgaard continued his “quest to find things to program” which brought him back to artificial intelligence. This time, however, he had more experience with math and could better understand the information provided to him.
He managed to create an AI that determined if a color looks better on a white or black background. The AI he made was very simple since he hadn’t completed precalculus yet, but it was still an improvement from his first exercise with AI.
Kollgaard continued his “quest” afterwards, which (right before his sophomore year) led him away from AI and towards game development. He, along with his friend and his older brother, developed a simple game for Android called Graze. This was a new world for Kollgaard as the skills required to make a game are very different from those required to make AI. Developing a game was fun and rewarding for him, but his quest wasn’t over.
The Problem
Between the spring of his sophomore year and the fall of his junior year Kollgaard worked at the Rita’s in Nazareth. Here he took orders, made orders, and served orders. He was somewhat of a trackstar, so he went to work later in the day. This forced him to watch as gallons of Italian ice were thrown away at the end of every day.
This waste is created for two reasons: the ice has to be made every morning and expires after only a couple of days. Jay Calandra, the owner of the Rita’s in Nazareth for over 8 years described the life cycle of ice: “Make it today, sell it today and tomorrow.”
After the second day, the ice has to be thrown away. This issue of getting rid of excess product at the end of the day isn’t unique to Rita’s, but it is unique how so much of their main product is thrown away.
Before the start of every work day, batches of ice have to be made. What is made is what is sold, so if too little ice is made then no one can buy it after it is sold out. Too much ice hurts business just as much since the mix for the ice has to be bought.
“[Rita’s] has to sell about one-third to one-fourth of what’s made to break even,” said Calandra. By selling a proprietary mix for ices to their stores, Rita’s incentivises owners to produce minimal waste. Unfortunately figuring out how much ice to make every day is a guessing game and an estimated 15%-20% of all that is produced is waste.
The Solution
Kollgaard came up with a clever solution to the guessing game. The amount of ice that people buy day-to-day may vary depending on weather, special occasions, seasons, work, etc. Now, for each individual person guessing what they may buy on a day-to-day basis would be very inaccurate, but with enough samples randomness can be modeled. Kollgaard realized that by using data from previous years it is possible to calculate how much of each ice is likely to be sold.
This may seem impossible due to confounding variables (unaccountable situations) such as people buying more when they randomly get a promotion or buying less after a break-up. These outlying events would mostly only affect a singular person, and since it would occur randomly they can be ignored.
Big events that could lead to a spike in sales such as school ending, weekends, or holidays could be accounted for by using the date to help calculate the amount of ice sold. Kollgaard described the previous day's orders, weather, and the date as having the most influence over what to produce..
There is one major flaw in Kollgaard’s plan: he needs a lot of data. Calculating the amount of ice consumed each day for a year would not provide enough data, so years of tracking orders would be required. At other businesses such as Red Robin or McDonalds this would be easy since they keep track of exact orders.
Unfortunately, at Rita’s the only thing kept track of on a receipt is the size and type of an order: not the flavor. This would be useful if each flavor was bought equally, but certain flavors are loved more than others. Knowing to make 50 gallons of ice isn’t really meaningful when you have 18 to 22 flavors to split across all at different, unknown proportions.
Keeping track of each flavor sold isn’t nearly as easy as it may seem. Relying on multiple people to keep careful records over several years does not always have the best results. If a lot of people come to Ritas at once then writing down what was ordered or keeping tally would become an obstacle in the way of making orders.
Requiring the flavor be recorded by the register would also be inefficient as it would add extra steps to placing an order. The flavors sold also change everyday and throughout the day, so each of the 81 unique flavors would have to be an option on the register. This would bloat any spreadsheet or interface used to record what is sold.
Kollgaard found a way to circumvent human error and inefficiencies. By placing a camera above where the ices are served, it is theoretically possible to track what orders were placed.
Collecting Data
The ability for a computer to detect information in an image is called computer vision. Right now in computer science, Convolution Neural Networks (CNNs) are the most popular way to give computers vision.
A neural network is a computer algorithm that attempts to mimic the nerves inside a human brain. It does this by taking an input (a picture of the ice in this case) and then using that information to activate different nodes. Nodes are like connection points and simulate neurons inside the brain. When one activates, it can activate many others.
What nodes are activated is decided by previously activated nodes and weights that a computer can calculate through a process of trial and error. A convolutional neural network is specifically designed to deal with detecting shapes, colors, outlines, etc. in an image. It achieves this by breaking an image into smaller pieces that it can manage easier. Filters are then applied to these smaller pieces to find certain shapes, outlines, etc.
Kollgaard is working on a CNN that would connect to a camera placed above where ice is served. Video taken from the camera would connect to the neural network so that it can discern the size of the cup and the flavor of the ice. While there are 81 possible flavors, there are only 18-20 available on any given day. Each flavor has a slightly different color which means his CNN should be able to discern between extremely similar colored flavors such as cherry and swedish fish,
Kollgaard said he is making his CNN the hardest way possible as he is attempting to make it from scratch. Templates of neural networks called frameworks are created to allow anyone to make a neural network. Kollgaard has opted out of using one so that he can learn more about AI.
Kollgaard came up with a clever solution to the guessing game. The amount of ice that people buy day-to-day may vary depending on weather, special occasions, seasons, work, etc. Now, for each individual person guessing what they may buy on a day-to-day basis would be very inaccurate, but with enough samples randomness can be modeled. Kollgaard realized that by using data from previous years it is possible to calculate how much of each ice is likely to be sold.
This may seem impossible due to confounding variables (unaccountable situations) such as people buying more when they randomly get a promotion or buying less after a break-up. These outlying events would mostly only affect a singular person, and since it would occur randomly they can be ignored.
Big events that could lead to a spike in sales such as school ending, weekends, or holidays could be accounted for by using the date to help calculate the amount of ice sold. Kollgaard described the previous day's orders, weather, and the date as having the most influence over what to produce..
There is one major flaw in Kollgaard’s plan: he needs a lot of data. Calculating the amount of ice consumed each day for a year would not provide enough data, so years of tracking orders would be required. At other businesses such as Red Robin or McDonalds this would be easy since they keep track of exact orders.
Unfortunately, at Rita’s the only thing kept track of on a receipt is the size and type of an order: not the flavor. This would be useful if each flavor was bought equally, but certain flavors are loved more than others. Knowing to make 50 gallons of ice isn’t really meaningful when you have 18 to 22 flavors to split across all at different, unknown proportions.
Keeping track of each flavor sold isn’t nearly as easy as it may seem. Relying on multiple people to keep careful records over several years does not always have the best results. If a lot of people come to Ritas at once then writing down what was ordered or keeping tally would become an obstacle in the way of making orders.
Requiring the flavor be recorded by the register would also be inefficient as it would add extra steps to placing an order. The flavors sold also change everyday and throughout the day, so each of the 81 unique flavors would have to be an option on the register. This would bloat any spreadsheet or interface used to record what is sold.
Kollgaard found a way to circumvent human error and inefficiencies. By placing a camera above where the ices are served, it is theoretically possible to track what orders were placed.
Collecting Data
The ability for a computer to detect information in an image is called computer vision. Right now in computer science, Convolution Neural Networks (CNNs) are the most popular way to give computers vision.
A neural network is a computer algorithm that attempts to mimic the nerves inside a human brain. It does this by taking an input (a picture of the ice in this case) and then using that information to activate different nodes. Nodes are like connection points and simulate neurons inside the brain. When one activates, it can activate many others.
What nodes are activated is decided by previously activated nodes and weights that a computer can calculate through a process of trial and error. A convolutional neural network is specifically designed to deal with detecting shapes, colors, outlines, etc. in an image. It achieves this by breaking an image into smaller pieces that it can manage easier. Filters are then applied to these smaller pieces to find certain shapes, outlines, etc.
Kollgaard is working on a CNN that would connect to a camera placed above where ice is served. Video taken from the camera would connect to the neural network so that it can discern the size of the cup and the flavor of the ice. While there are 81 possible flavors, there are only 18-20 available on any given day. Each flavor has a slightly different color which means his CNN should be able to discern between extremely similar colored flavors such as cherry and swedish fish,
Kollgaard said he is making his CNN the hardest way possible as he is attempting to make it from scratch. Templates of neural networks called frameworks are created to allow anyone to make a neural network. Kollgaard has opted out of using one so that he can learn more about AI.
Issues
In a perfect world, some issues would still exist with Kollgaard’s method and some waste would still be produced.
Even if every bit of ice that is produced and made for sale is sold, there would still be an estimated 5% of ice that gets stuck on the scoopers, stuck on the machine, or wasted by the customers. In order for no ice to go to waste then it would all have to be perfectly mixed, poured, and scooped. This is impossible in the real world, so Kollgaard is only attempting to mitigate the amount of ice being thrown away.
After collecting all the data, Kollgaard won’t be done with preventing waste. He will still need to find a way to transfer the data into an algorithm to determine how much ice to make and create a way to input data daily for the algorithm.
Getting input will be easy, but actually putting together an algorithm to accurately estimate how much ice to produce with a margin of error will be difficult.
“Modeling a real world event like [selling Italian ice] is no easy task,” said expert business analyst and data scientist Daniel Dewey. “Sure, you can use last year’s sales as a baseline, but externalities like unpredictable weather patterns, market competition, and changing [consumer] tastes layer complex variances in year on year sales.”
Essentially, in order to stay accurate Kollgaard would constantly need to be collecting new data. Kollgaard will likely plan for this and add data as it’s input into the algorithm, but the algorithm would still need to be complicated in order to account for all possible variances.
Accounting for variances would lead to forming a range of possible ices sold which would cause at least a little bit of waste in order to maximize profit. This is why even in a perfected world at least 5% of product would be expected to be wasted. While not ideal, even 7% would be better than the 15% - 20% that is currently expected. Kollgaard is making strides in reducing wastes produced by Ritas.
In a perfect world, some issues would still exist with Kollgaard’s method and some waste would still be produced.
Even if every bit of ice that is produced and made for sale is sold, there would still be an estimated 5% of ice that gets stuck on the scoopers, stuck on the machine, or wasted by the customers. In order for no ice to go to waste then it would all have to be perfectly mixed, poured, and scooped. This is impossible in the real world, so Kollgaard is only attempting to mitigate the amount of ice being thrown away.
After collecting all the data, Kollgaard won’t be done with preventing waste. He will still need to find a way to transfer the data into an algorithm to determine how much ice to make and create a way to input data daily for the algorithm.
Getting input will be easy, but actually putting together an algorithm to accurately estimate how much ice to produce with a margin of error will be difficult.
“Modeling a real world event like [selling Italian ice] is no easy task,” said expert business analyst and data scientist Daniel Dewey. “Sure, you can use last year’s sales as a baseline, but externalities like unpredictable weather patterns, market competition, and changing [consumer] tastes layer complex variances in year on year sales.”
Essentially, in order to stay accurate Kollgaard would constantly need to be collecting new data. Kollgaard will likely plan for this and add data as it’s input into the algorithm, but the algorithm would still need to be complicated in order to account for all possible variances.
Accounting for variances would lead to forming a range of possible ices sold which would cause at least a little bit of waste in order to maximize profit. This is why even in a perfected world at least 5% of product would be expected to be wasted. While not ideal, even 7% would be better than the 15% - 20% that is currently expected. Kollgaard is making strides in reducing wastes produced by Ritas.
