Warehouse route optimization

​Reinforcement Learning approach applied to warehouse route optimization task. The main goal was to build a simulation and to train a neural network using this simulation as an example.
​

The warehouse was presented in the form of a two-dimensional map on with various shelves filled with boxes. The breakthrough idea was to train a neural network not only using the predefined data but building the virtual environment that generates such kind of data on the fly and allows to train an agent in a very flexible way adjustable to different circumstances. On this project, we used an advanced version of machine learning.
​

According to the created rules the agent had to learn to play this warehouse game as efficiently as possible. This project is very similar to one that was performed by Amazon. Our team created the warehouse simulation engine from scratch (simulation objects, logic, screen rendering, etc.). Next step was to wrap that simulation into the environment via Gim framework (action space, observation space, reward function, etc.) Preparation and training of the Q-learning model were done via SciPy and NumPy. Preparation and training of the Deep-Q-Network — using Keras and Keras-RL frameworks.