Hierarchy through composition with multitask LMDPs

Conference item

Hierarchical architectures are critical to the scalability of reinforcement learning methods. Most current hierarchical frameworks execute actions serially, with macro-actions comprising sequences of primitive actions. We propose a novel alternative to these control hierarchies based on concurrent execution of many actions in parallel. Our scheme exploits the guaranteed concurrent compositionality provided by the linearly solvable Markov decision process (LMDP) framework, which naturally enables a learning agent to draw on several macro-actions simultaneously to solve new tasks. We introduce the Multitask LMDP module, which maintains a parallel distributed representation of tasks and may be stacked to form deep hierarchies abstracted in space and time.

Authors: Saxe, A

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: Journal of Machine Learning Research
• Host title: Proceedings of the 34 th International Conference on Machine Learning
• Journal: Proceedings of Machine Learning Research
• Volume: 70
• Issue: 2017
• Publication date: 2017-08-11
• Acceptance date: 2017-03-01
• Event title: International Conference on Machine Learning
• Event location: Sydney, Australia
• Event website: https://icml.cc/Conferences/2017
• Event start date: 2017-08-07
• Event end date: 2017-08-11
• ISSN: 2640-3498
• Language: English
• Keywords: FFR