Nguyen Minh Quang

[paper][code][poster]

Value-based reinforcement learning is the current State-Of-The-Art due to high sampling efficiency. However, our theoretical and empirical studies show evidence that it suffers from low exploitation in early training period and bias sensitiveness.  To address these issues, we propose to augment the decision-making process with hypothesis, a weak form of environment description.  Our approach relies on prompting the learning agent with accurate hypotheses, and designing a ready-to-adapt policy through incremental learning.  We propose the ALH algorithm and monitor it on detailed analyses on a typical learning scheme and a diverse set of Mujoco benchmarks.  Our algorithm produces a significant improvement over value-based learning algorithms and other strong baselines.

We propose a novel vertical scaling approach. In this approach, we conditionally factorize the multi-document output probability by lower-complexity components.  Specifically, these components are estimated by estimators optimized for single-doc data.  Unlike the full-attention approach, vertical scaling has complexity that scales linearly with the number of single documents, making it more efficient for long documents or large numbers of documents. To further enhance the efficiency and effectiveness of our approach, we introduce the Multi-Channel Attention architecture.  This architecture enables us to fully utilize BART's single-doc pre-optimized parameters, while does not require re-optimization, leading to a zero-cost transition. Our approach maintains promising accuracy and computing efficiency.