- paper link: https://arxiv.org/abs/2005.01643
1 min read
Offline Reinforcement Learning - Tutorial, Review, and Perspectives on Open Problems The Optimizer’s Curse - Skepticism and Postdecision Surprise in Dec...
Introduction 추천 시스템에서 발생하는 loop 는 bias 를 일으킴 (1) collect data, (2) train model, (3) deploy model loop off-policy learning 에서 biased data 로 부터 unbiased learning 을 지향하는 방법에는...
Paper link arxiv.org/abs/1812.07127 .
References paper link: arxiv.org/pdf/2109.10665.pdf Reinforcement Learning based Recommender Systems: A Survey .
Discussion Q-learning based offline 학습 방식이고, 모델도 무겁고.
Abstract 뉴스 추천을 위한 딥러닝 기반의 강화 학습 프레임워크를 제안한다. news feature 들과 user 의 preferences 의 변동성 (dynamic) 을 설명하는 것은 상당히 어렵다.
Empirical Evaluation: Live Experiments YouTube 에 SARSA-TS 알고리즘을 실험 candidate -> ranker 를 거치게 되는데, ranker 의 scoring 함수에서 사용하는 myopic(근시안적) engagement 측정값을 LTV estimate 로 변경함...
Abstract S-MDP 문제를 다룬 paper 이다.
Abstract 해결하려는 문제: 강화학습에서의 효율적인 exploration Randomized value functions offer a promising approach to efficient exploration with generalization, but existing algorithms are not...
paper Link: arxiv.org/pdf/1205.2606.pdf KWIK Linear Regression KWIK (Knows What It Knows) is a framework for studying supervised learning algorithms and was designed to unify...
RL methods work by maximizing the expected return of a policy.