Session 14 - Healthcare: Exploring Task Decomposition with Ensemble Approach using Reinforcement Learning
Presentation Type
Oral
Student
Yes
Track
Methodology
Abstract
Keywords: Reinforcement Learning, Ensemble Method, Task decomposition.
In the context of machine learning, Reinforcement Learning (RL) hardly requires an introduction and the ensemble method is an effective approach that can improve the performance of a model. In a supervised learning environment, it has been shown that an ensemble of individual learners is capable of decomposing the input space of a classification task. Our research is intended to ascertain whether a model consisting of an ensemble of Q-learners can decompose a specific control task into distinct components and train itself to designate individual learners for a specific component without any additional information or supervision from the human trainers.
A number of artificial intelligence concepts are based on human behavior and how we learn. Task decomposition also has a relevance to human behavior. For a group of people working at a problem, it often happens that the problem is broken down into discrete smaller jobs which are assigned to the individuals in the group. We intend to investigate if this concept can be translated into RL domain by using an ensemble of Q-learners.
While the relevant reviewed papers indicated enhancement of RL performance with ensemble approach, it does not offer any prospect that might be used for task decomposition, without any additional task parameters. With our research we want to utilize this unexplored potential and look into the possibility of an ensemble of learners decomposing the input space of a control task.
Start Date
2-11-2020 3:30 PM
End Date
2-11-2020 4:30 PM
Session 14 - Healthcare: Exploring Task Decomposition with Ensemble Approach using Reinforcement Learning
Campanile & Hobo Day Gallery (A & B)
Keywords: Reinforcement Learning, Ensemble Method, Task decomposition.
In the context of machine learning, Reinforcement Learning (RL) hardly requires an introduction and the ensemble method is an effective approach that can improve the performance of a model. In a supervised learning environment, it has been shown that an ensemble of individual learners is capable of decomposing the input space of a classification task. Our research is intended to ascertain whether a model consisting of an ensemble of Q-learners can decompose a specific control task into distinct components and train itself to designate individual learners for a specific component without any additional information or supervision from the human trainers.
A number of artificial intelligence concepts are based on human behavior and how we learn. Task decomposition also has a relevance to human behavior. For a group of people working at a problem, it often happens that the problem is broken down into discrete smaller jobs which are assigned to the individuals in the group. We intend to investigate if this concept can be translated into RL domain by using an ensemble of Q-learners.
While the relevant reviewed papers indicated enhancement of RL performance with ensemble approach, it does not offer any prospect that might be used for task decomposition, without any additional task parameters. With our research we want to utilize this unexplored potential and look into the possibility of an ensemble of learners decomposing the input space of a control task.