Full Record

New Search | Similar Records

Author
Title Learning competitive ensemble of information-constrained primitives
URL
Publication Date
Date Accessioned
University/Publisher Université de Montréal
Subjects/Keywords Reinforcement Learning; Hierarchical Reinforcement Learning; Information Bottleneck; Compositionality; Modular network; Apprentissage par renforcement; Apprentissage par renforcement hiérarchique; Goulot d'étranglement de l'information; Compositionnalité; Réseaux modulaires; Applied Sciences - Artificial Intelligence / Sciences appliqués et technologie - Intelligence artificielle (UMI : 0800)
Contributors Bengio, Yoshua (advisor)
Rights Unrestricted
Country of Publication ca
Record ID handle:1866/22537
Repository montreal
Date Indexed 2020-08-12
Issued Date 2019-10-30 00:00:00

Sample Search Hits | Sample Images

…decentralized manner. We review the information theory literature in section 1.4. The general idea of training modular networks goes beyond Hierarchical Reinforcement Learning and is studied under the paradigm of Neural Modular Networks, which we review in…

Learning LSTM Long-Short Term Memory MDP Markov Decision Process NMN Neural Module Network PVF Proto-Value Functions RL Reinforcement Learning RNN Recurrent Neural Networks x To my family ! xi Acknowledgments I thank the almighty god for giving me a…

…and not just about research. I thank my undergraduate research advisor Dr Dhaval Patel for accepting me as his student and always encouraging me to pursue research. xii 1 Introduction We want to develop reinforcement learning algorithms that can…

…quickly adapt to new tasks by obtaining a structured decomposition of the learning agent’s behavior. This problem is well studied under the paradigm of Hierarchical Reinforcement Learning. We review this body of literature in section 1.3 and describe the…

…behind this decomposition is as follows: When we have a standard policy, the policy needs to specialize in the entire state space. In the case of Hierarchical Reinforcement Learning, the primitives have to specialize in only a part of the state space…

…environment to collect training experience and instead of being given explicit supervision in terms of what input maps to what output, the system has to learn based on the “rewards” it gets from the environment. This is the well known reinforcement learning

…paradigm [Sutton and Barto] (summarized in section 1.2). Deep Learning and Reinforcement Learning based techniques have transformed many application domains - image classification, image generation, machine translation, question…

…at this discrepancy from the lens of compositionality (section 1.5) and look at architecture designs like hierarchical reinforcement learning (section 1.3) and neural module networks (section 1.9) that can incorporate the…

.