University of Twente Student Theses
Towards an Optimal Initial Sparsity Distribution for Automatic Noise Filtering in Deep Reinforcement Learning
Dragomir, Ecaterina (2023) Towards an Optimal Initial Sparsity Distribution for Automatic Noise Filtering in Deep Reinforcement Learning.
PDF
2MB |
Abstract: | Efficiently learning how to select relevant input features remains a daunting task in the development of artificial intelligence models, especially given high noise levels (80\% or more). However, when autonomous systems tauntingly spring into our daily, noisy lives, the importance of automatically and accurately detecting relevant features cannot be ignored. Considerable progress in this direction has been made in the last few years, by using partially sparse (i.e. the output layer is dense) Deep Reinforcement Learning models, such as Automatic Noise Filtering (ANF) \cite{Grooten23}, which can outperform previously existing networks in environments augmented with up to 98\% Gaussian noise. This discovery begs the question of what an optimal sparsity distribution looks like, given a noisy medium. Thus, the present research aims to answer this question, by analysing the effects that various sparsity distributions - inspired from fields such as artificial intelligence, cognitive neuroscience and mathematics - have on the learning efficacy and speed of ANF \cite{Grooten23} in highly noisy environments. The results validate the decision to maintain a dense output layer, as well as support the proposal of a series of novel sparsity distributions, among which an inverse Erdős–Rényi model. |
Item Type: | Essay (Bachelor) |
Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
Programme: | Business & IT BSc (56066) |
Awards: | Best paper award |
Link to this item: | https://purl.utwente.nl/essays/95985 |
Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Repository Staff Only: item control page