Setup and what AI is

At its core, artificial intelligence means any technology that simulates human functions. Most of the time, instead of following strict rules, you show it many examples and let it find the rules on its own, through various algorithms. It isn't magic: ChatGPT doesn't "think". The part about learning from examples is called machine learning, and it's almost everything you do at the olympiad.

Some terminology is worth mentioning too. Data science means getting useful information out of data, using mathematics (statistics above all) and computer science. Mathematics is the foundation and shows up in absolutely everything we do, but you don't have to learn it in depth to solve problems or build projects. Some classic computer science algorithms are used as well, BFS and DFS for example, in certain areas of AI. Machine learning, mentioned above, splits into several categories of its own: supervised learning, unsupervised learning, reinforcement learning and, sometimes, self-supervised.

A few examples help. To tell a spam email from a normal one with hand-written rules, you'd have to write hundreds of conditions of the form "if it contains the word X". With machine learning, you give it a few thousand emails already marked as spam or not, and the model learns by itself which combinations of words predict spam. A few classic AI tasks include, but aren't limited to:

Overfitting, underfitting, bias and variance matter too. A model that has memorized the training data, noise included, does almost perfectly on what it has seen and poorly on new data: that's overfitting. At the opposite end, a model too simple for the problem doesn't even capture the real pattern and gets things wrong everywhere, on training and on test alike: that's underfitting.

The two are explained by bias and variance. Bias is the error that comes from assumptions that are too simple, like fitting a straight line through data that actually describes a curve. Variance is sensitivity to the exact data it was trained on: change the training set slightly and the model gives completely different predictions. High bias means underfitting, high variance means overfitting, and lowering one usually raises the other. Everything you do at the olympiad, from how complex a model you pick to how long you train it, is really the search for the balance point between them.

Python is the standard language for writing artificial intelligence code. Not because it's the fastest, but because it's heavily abstracted and already has the libraries you want, written for you, sometimes in faster languages: NumPy for numerical computing, Pandas for tables, scikit-learn for classic models, PyTorch for neural networks.

A library is code written by someone else that you import and use. In AI tasks the goal isn't to memorize how algorithms or models are written, but to understand them, to know when each one applies, and to adapt them to the problem at hand alongside other methods.

Before you write any code you need three things installed: Python, pip (which comes with it automatically) and Jupyter Lab. Jupyter Lab, VS Code and Google Colab are environments where you can write Python notebooks, which let you write code in pieces and run it one piece at a time.

Pip ships with Python from version 3.4 on, so you don't install it separately. To confirm both work, open a terminal (Command Prompt or PowerShell on Windows, Terminal on macOS and Linux) and run:

On Windows, if python isn't recognized, try py --version; if pip doesn't answer either, try pip3 or python -m pip --version.

With Python installed, move on to the working environment itself. A virtual environment is a separate box for one project's libraries, so they don't clash with others installed globally. You create it once per project and always work inside it.

At the olympiad you work on the contest centre's computers, with Windows, JupyterLab, Python 3.11 and Firefox. You can't bring notes or books into the room, and online you can only reach the contest platform and the official documentation of the language and the allowed packages, for example docs.python.org, numpy.org/doc, pandas.pydata.org/docs or scikit-learn.org. The exact list is announced by the committee.

That's why it's worth getting used to searching the documentation now, not Google. You also get help right inside the notebook: type a function's name followed by ?, for example pd.read_csv?, or press Shift+Tab inside its parentheses to see which arguments it takes.