Knowledge- and Gradient-Guided Reinforcement Learning for Parametrized Action Markov Decision Processes
Jonas Ehrhardt, René Heesch, Oliver Niggemann
Read on arXiv →
Key claim
KGRL improves sample efficiency in PAMDPs using domain knowledge.
In plain English
Reinforcement Learning often struggles with sample efficiency, especially in complex decision-making scenarios like PAMDPs. Current methods typically rely on one-shot estimators, which can lead to inefficiencies. This paper introduces KGRL, which leverages existing domain knowledge to refine decision-making and improve training efficiency. Builders might care because KGRL not only enhances performance but also provides explanations for its decisions, making it easier to understand and trust the agent's behavior.
Introduces a novel approach combining symbolic reasoning with RL for PAMDPs.
Demonstrates improved performance over state-of-the-art baselines with clear methodology.
Deep reliability assessment
The methodology supports the claim that integrating domain knowledge into RL can improve sample efficiency and performance, but the reliance on specific domain knowledge may limit generalizability across diverse environments.
Reproducibility
Yes, the paper provides open source code for reproducing the results: https://github.com/j-ehrhardt/kgrl.
Key figure
Figure 3 shows the learning curves of KGRL and baseline algorithms across different evaluation domains, illustrating training progress and mean episodic return.