A video game I developed to capture an amazing trip across the Tunisian Sahara, honoring the beautiful landscapes and rich heritage of the region.

Built an agent that uses a swarm of solvers to tackle ARC-AGI problems, with a design backed by a small bank of reusable Domain-Specific Language (DSL) primitives.

Developed a Proximal Policy Optimization (PPO) approach for solving autonomous racing tasks in the AWS DeepRacer simulation environment. The approach employs a convolutional neural network architecture designed to process multi-modal sensor inputs, including stereo camera images and LIDAR distance measurements.

Implemented and trained a Multi-Agent Proximal Policy Optimization (MAPPO) approach for cooperative cooking tasks in the Overcooked simulation environment, using centralized training with decentralized execution — agents learn independently while sharing the same environment.

Studied the REINFORCE policy-gradient algorithm on the Lunar Lander environment to better understand policy-gradient methods for continuous control. Implemented REINFORCE with continuous actions in PyTorch and tuned its hyperparameters to see which ones matter most for performance.

Implemented and trained several reinforcement learning algorithms (Value Iteration, Policy Iteration, SARSA, and Q-learning) and compared their performance on a traffic-management control task with the objective of minimizing car congestion.

Apply Policy Gradient approaches to teach an agent to play the game Pong from the PyGame Learning Environment.

Try out different architectures of neural nets using PyTorch in order to predict a comparison of two handwritten digits.

Implementing modern deep learning models can hardly be done without a proper framework to minimize code duplication and maximise ease of use as well as good structure. In this project, we implemented our own mini-framework and compared our performance with pyTorch’s NN library, from which it was inspired.

Use temporal information about Twitter hashtags to discover trending topics and potentially uncover world events as they occurred.

We study the problem of building an efficient recommender system. We are only given access to a subset of users and their ratings, and we aim to recommend new movies by predicting the missing ratings. To this end, we considered a collaborative-based filtering approach along with ensemble methods.

We explore and compare different supervised learning algorithms and how they deal with a data-set from CERN in the field of physics to predict the presence of the Higgs Boson.