Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Active Learning for Predicate Classification
The aim of this project is to develop a learning-based classification system that queries a user for labels on examples that are the most informative for improving the system’s prediction performance.
Keywords: Machine learning, user interaction, few-shot learning
Highlevel planning is used in robotics to decide on which actions a robot should take to achieve a given goal. A common way to model such planning problems is based on so-called predicates, which are binary state variables describing an environment (e.g. “inside”, “on”, “stacked”, “inserted”, …) [1].
To apply highlevel planning, classifiers for all relevant predicates are needed. However, in complex real-world environments, there are a large number of possible predicates that creative users could think of, and task their robot with achieving. Therefore, we would like to build a system that can serve as a classifier for predicates.
To make giving demonstrations to that system bearable, it should be trainable with as few data points as possible. To achieve this, we envision an interactive process [2, 3], during which the system generates scenarios that the user then labels. The scenarios should be generated so that the resulting labels are maximally informative for the system, allowing it to achieve good classification performance with very few user interactions.
If this sounds like an interesting challenge to you, we would be happy to hear from you.
[1] E. Karpas and D. Magazzeni, “Automated Planning for Robotics,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 3, pp. 1–23, 2019.
[2] Evensen, S., Ge, C., Choi, D., & Demiralp, Ç. (2020). Data Programming by Demonstration: A Framework for Interactively Learning Labeling Functions. arXiv preprint arXiv:2009.01444.
[3] Chernova, S., & Veloso, M. (2009). Interactive policy learning through confidence-based autonomy. Journal of Artificial Intelligence Research, 34, 1-25.
Highlevel planning is used in robotics to decide on which actions a robot should take to achieve a given goal. A common way to model such planning problems is based on so-called predicates, which are binary state variables describing an environment (e.g. “inside”, “on”, “stacked”, “inserted”, …) [1].
To apply highlevel planning, classifiers for all relevant predicates are needed. However, in complex real-world environments, there are a large number of possible predicates that creative users could think of, and task their robot with achieving. Therefore, we would like to build a system that can serve as a classifier for predicates.
To make giving demonstrations to that system bearable, it should be trainable with as few data points as possible. To achieve this, we envision an interactive process [2, 3], during which the system generates scenarios that the user then labels. The scenarios should be generated so that the resulting labels are maximally informative for the system, allowing it to achieve good classification performance with very few user interactions.
If this sounds like an interesting challenge to you, we would be happy to hear from you.
[1] E. Karpas and D. Magazzeni, “Automated Planning for Robotics,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 3, pp. 1–23, 2019. [2] Evensen, S., Ge, C., Choi, D., & Demiralp, Ç. (2020). Data Programming by Demonstration: A Framework for Interactively Learning Labeling Functions. arXiv preprint arXiv:2009.01444. [3] Chernova, S., & Veloso, M. (2009). Interactive policy learning through confidence-based autonomy. Journal of Artificial Intelligence Research, 34, 1-25.
- Literature review
- Design and implement classifier and generator architectures
- Evaluate system
- Literature review - Design and implement classifier and generator architectures - Evaluate system
- Highly motivated and independent student
- Interest in machine learning
- Good programming skills in Python
- Experience with the following is a plus: PyTorch, Git
- Highly motivated and independent student - Interest in machine learning - Good programming skills in Python - Experience with the following is a plus: PyTorch, Git
Daniel Dugas (daniel.dugas@mavt.ethz.ch)
Julian Förster (julian.foerster@mavt.ethz.ch)
Daniel Dugas (daniel.dugas@mavt.ethz.ch) Julian Förster (julian.foerster@mavt.ethz.ch)