Core Concepts
Implementing novelty detection within world model RL agents is crucial for protecting agent performance and reliability.
Abstract
Abstract:
Novelty detection is essential to maintain agent performance.
Proposed bounding approaches for incorporating novelty detection.
Introduction:
Reinforcement learning using world models has seen recent success.
Addressing the under-explored area of novelty detection in RL.
Data Extraction:
"arXiv:2310.08731v2 [cs.AI] 22 Mar 2024"
Related Work:
Novelty detection applications and challenges in RL.
Background:
Partially observable Markov Decision Processes explained.
DreamerV2 World Model:
Components and functions of the DreamerV2 model.
Bayesian Surprise:
Theory behind measuring surprise in observations.
Latent-based Detection:
Bound formulation without hyperparameters for novelty detection.
Experiments:
Evaluation of proposed methods on MiniGrid environments.
Baselines:
Comparison with RIQN, CMTRE, and PP-MARE methods.
Metrics & Results:
Average delay error, false alarms, precision, recall, accuracy, and AUC results.
Stats
"arXiv:2310.08731v2 [cs.AI] 22 Mar 2024"