Hi there! And welcome to the 256th day of the year. It is also known as "Programmer's Day" and it's a very special holiday for us. Although our latest games are not programming sims per se, they do revolve around coding. Furthermore, we believe that coders are the heroes of the XXI century, those capable of changing the world.
In other words, if you are not into coding yet, you better start catching up! And Learning Factory might be just the game for you to start with.
If you have played the alpha, you might have noticed that Learning Factory is not just about building a factory: it's about understanding and predicting the wishes of your customers, the cats. Although the concept of cats buying stuff at some factory might seem a little abstract if not absurd (to put it mildly), it is based on real-life research methods. In this post we'll try and shed some light on how we (and you!) will be using actual, real-life machine learning in this game.
Ivan Kapranov, our team's own lead coder, holds a bachelor's degree in machine learning. Even more so, he's teaching it to high school students. To commemorate Programmer's Day, we've asked Ivan to shed some light on the learning part behind the Learning Factory.
DISCLAIMER: We do understand that machine learning and coding are closely related, yet they are not the same things. Our game is specifically about the latter. If you want to learn more about coding by playing video games, please feel free to check out respective games, such as Human Resource Machine or any of Zachtronics' games. They're awesome!
"We're planning to introduce several algorithms of machine learning into the game. The one I'm working on right now is linear regression that will help predicting prices for goods in the game's stores. Each of them is selling several items which can be purchased by cats. Naturally, the store's mission is to maximize the profit. Linear regression can help it a great deal. A player's main goal is to provide data (in general, the more -- the better, but it's not really that simple) to be processed by the algorithm so it could calculate the optimal price for each type of product within the data provided. E.g. if the data is of good quality, the price will be close enough to optimum, if the data is bad -- you'll get apples vs. oranges.
What kinds of features will we have in the game? For starters there will be a rather simple regression based on a few simple features, but the price predicted by such an algorithm will be far from optimum, because each cat is different and is willing to pay a certain price. Therefore a player will need to add additional features to the algorithm. We will also have sales, special offers and such… Also we plan to make different kinds of cats. Some of them might be first time customers, some would be returning. Some of them might be holding loyalty cards. Or maybe discount coupons! In order to make a factory efficient and profitable, one would have to take all these factors (and more!) into consideration.
This is where linear regression does its magic. In theory, it's possible to calculate the optimal price by hand, especially if you don't mind spending a couple of days collecting and analysing data. And even after that you'd have to change the price every couple of minutes, because the circumstances are constantly changing. Is there a new president elected on Mars? Or maybe there's a problem with the supply chain of a famous cheese from Venus? Every global or local event might change the demand for goods. If you have a well tuned algorithm, it will adapt to these changes real fast by analyzing new kinds of data. (Or so we expect, because the data we're using is really simple, but we still are in the early stages of development)
Machine learning helps to take all the factors into consideration. But before that would become possible, a player would have to prepare raw data for the algorithm. Actually, this is what 95% of working with machine learning in production is all about. In Learning Factory, we are trying to make this process as graphic and illustrative as possible -- hence the factory building concept. Different kinds of buildings would collect different kinds of data, which then flows to data cluster cores where the mathematical magic is happening. Say, a player wants to build a model which would acknowledge the price of goods, the type of store and something less obvious, like, cats' ears shape. They would have to streamline the data to a cluster core, push a button and let the algorithm calculate the optimum price within the dataset.
This concludes the first iteration of machine learning, but then there's an option to go deeper and actually experiment. This is where a player will be working with regularizations that help algorithms learn better, to put it simple. They would have to learn about weights of features and how they influence the price prediction quality.
And then a player can go even deeper. We plan to introduce special kinds of buildings that will allow players to experiment with data -- multiply additional features, for example. Or exponentiation! Say, a player wants to check out what happens if an algorithm calculates the squared price of a type of goods. Maybe this way it will be more efficient. Or maybe not -- it is an experiment, after all! Another way to improve the quality of prediction is to input additional features into the model, which would also require building a certain facility in a factory.
The mission of Learning Factory as we see it is to illustrate linear regression (and maybe other kinds of machine learning, stay tuned for updates!) in a fun way. If you apply conservative approach to learning, you'd have to listen to several hours of lectures, then install Python and certain libraries and spend another handful of hours on coding… That is, if you have the necessary knowledge in maths and know how to write a code. The game allows you to skip all the hard stuff and still be able to grasp the general context. We want to make Learning Factory as flexible for the players as possible: you don't have to go too deep into machine learning to make your factory profitable, but those who choose to do so will discover a whole new layer of gameplay to experiment with. And if you want to dive deeper into machine learning context… well, let's just say, we have some ideas on how to bring the game closer to real life! But we'll talk about it when it's time.
We are bold enough to believe that a good video game can be a perfect supplementary material to lectures on YouTube or Coursera. And this time, things are real! Our previous game while True: learn() was merely simulating machine learning at most times, but Learning Factory will actually feature living and breathing ML algorithms for a player to experiment with. It won't make you into a real data scientist, but it can be your first step to becoming one!
In a traditional learning model, a contact with audience is very important: old school lecturing implies face to face contact, and that does not really transit into online mode. If one's trying to learn something online, it requires a very high level of inner motivation. Video games, on the other hand, are self motivating, because they're fun! And if a player absorbs some 10% of new information just by playing a game, I believe our job as developers and educators is well done!"
[h2]More Engineers Required[/h2]
We want YOU at KOTOVOD! Your participation in our Alpha program will help us create a more bright and enlightened future for cats and humans! Please use the information below to join our ranks:
To play the Alpha, please join our
Discord server or sign up for
Email listAnd of course, don't forget to
add Learning Factory to your Wishlist (check out the button on the right of this page!)
Learning Factory's Public Roadmap
