Harnisch, Richard (2026) On The Existence of Double Descent in Reinforcement Learning. Bachelor's Thesis, Artificial Intelligence.
|
Text
thesis.pdf Download (3MB) | Preview |
Abstract
Double Descent (DD) is a test risk phenomenon in which a deep model's test risk worsens around an interpolation threshold (overfitting), but then improves again when increasing training episodes, model size, or dataset size. We implement an empirical experiment to test whether this phenomenon can appear in Reinforcement Learning (RL). To this end, we train Deep Q-Network (DQN) and Trust-Region Policy Optimization (TRPO) agents on a family of seeded gridworld mazes and evaluate the agent on held-out maps. Across model sizes and training duration, we observe a generalization gap appearing between returns on train and held-out test maps, but no second descent regime in test risk upon increasing the size of the model as well as the amount of training episodes. The results suggest that in this context, increased capacity and training do not recover generalization and motivate further experiments with different algorithms, architectures, and environment types.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Sabatelli, M. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 01 Apr 2026 09:27 |
| Last Modified: | 01 Apr 2026 09:27 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37276 |
Actions (login required)
 |
View Item |