University of Denver
Collaborative Federated Learning for Robots in Heterogeneous Environments
Abstract
dc:description.abstractThis research investigates the performance of Federated Averaging (FedAvg) in simulated Federated Learning (FL) scenarios with varying degrees of environmental heterogeneity among robotic agents. The study explores the impact of data heterogeneity on both the convergence of FedAvg and the fairness of learning, with regard to consistency of performance across agents. Experiments were conducted with simulated robots trained to perform a target collection task, where a subset of agents encountered an unfamiliar environment. The results demonstrate that while FedAvg exhibits resilience to the introduction of new environmental data, it struggles to ensure both convergence and fairness in heterogeneous settings. Specifically, the original training environment maintains a dominant influence, while agents in novel environments experience inhibited learning. These findings highlight the challenges of applying FedAvg in lifelong learning scenarios with heterogeneous environments and unbalanced non-IID data, revealing a trade-off between maintaining performance in a majority environment and ensuring equitable learning across diverse environments.
Degree
thesis:*- Name thesis:degree_name
- M.S. in Computer Science
- Level thesis:degree_level
- Masters Thesis
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Schneider, Karlan
- Contributors dc:contributor
-
- Stephen Hutt
- Mohammad Mahoor
- Kerstin Haring
- Aaron Kraft
Subjects
dc:subject × 8Rights
dc:rights- Statement dc:rights
-
- <p>Copyright is held by the author. User is responsible for all copyright compliance.</p>
- Karlan Schneider
- Language dc:language
- English (eng)