<ul><li>The article discusses the importance of rule generalization for Reinforcement Learning (RL) agents to perform well on both seen and unseen entities.</li><li>It highlights the challenge of excessive generalization leading to false-positive results and the need to strike a balance.</li><li>A novel approach of dynamic rule generalization is proposed using WordNet's hypernym-hyponym relations.</li><li>The algorithm dynamically generates generalized rules based on information gain from positive and negative examples.</li><li>The study demonstrates the benefit of integrating symbolic and neural reasoning in RL agents for text-based games like TW-Cooking.</li><li>Experiments show that the EXPLORER agent outperforms neural-only agents and SOTA models like GATA and CBR on TW-Cooking and TWC games.</li><li>Different generalization settings in EXPLORER, such as Exhaustive Rule Generalization and IG-based generalization, are compared in the study.</li><li>Agents are trained with 100 episodes on TW-Cooking domain without pretraining advantage to boost performance.</li><li>The paper is available on arxiv under CC BY 4.0 DEED license.</li></ul>

Beyond Seen Worlds: EXPLORER’s Journey into Generalized Reasoning

Discover more