Marc Gröling

Examined a state-of-the-art data structure to maintain a dynamic set under lookup-queries. Implemented the theoretical construction proposed in the paper and conducted experiments on it, getting one step closer to make it usable for practice.

Created a waste tracking system with blockchain that uses a byzantine fault-tolerant consensus algorithm. A unique barcode is assigned to each electronic device and then added to the blockchain.

Built a system that can classify waste together with another student. Furthermore, a dataset of ~5000 images was collected from a local waste site. On this dataset, several image classification models were trained and used to classify waste. The work was presented in the Intelligent Systems Conference 2023.

Built a render engine to create high-definition images using ray-casting. The engine is built in C++ and covers lighting, textures, and a parallelised ray-casting procedure.

Created a reinforcement learning environment for the 3-sided Rubik's Cube in Python. The environment features a 3D visualisation tool that can render any object consisting of triangles using rasterization. However, it only runs on the CPU and is thus quite slow.

Trained an artificial neural network with soft Q imitation learning (SQIL), which uses a combination of supervised learning and reinforcement learning to generate training data. Hyperparameter optimisation was done and metrics to evaluate the models created.

Trained an LSTM neural network to predict the locomotion behaviour of elephant fish. Data (ray-casts and movement) was extracted from given recordings.The project was done as part of a software project in collaboration with two other people.