POKÉMON BATTLE SIMULATOR

While the theme chosen for this study may seen silly, we apply advanced machine learning techniques, which can also be applied to a range of applications such as sports competitions and market, sales, retail or customer stratification.

​

Pokémon is a famous Japanese game franchise, created by Nintendo, which later also expanded to other areas like trading card, manga, anime, and turned into global pop culture.

​

In this study we apply advanced neural network concepts like embedding  and shared layers and multiple inputs. We extract the concept of NLP models and n-grams and adapt them to our own problem.

​

In order to achieve the final model, we also perform data analysis, visualization and dimension reduction, with PCA and t-SNE techniques.

​

For our model we used the stats, types and moves features of Pokémon from 1st to 6th generation (around 700 Pokémon) and we train with 10,000 battles entries. After training, our model has an accuracy of 86% predicting winners of 8,000 battles, which contains also Pokémon from 7th and 8th generations (around 300 new Pokémon), previous unseen to the model.

​

The entire source code can be found in my repository.