We would like to invite submissions for the 1st Workshop on Goal
Specifications for Reinforcement Learning at the Federated AI Meeting 2018.
The submission deadline is May 1st 2018. Attached below is the call for
papers. We look forward to your submissions.
Paper submission opens: April 1st, 2018
Submission deadline: May 1st, 2018
Author notification: June 1st, 2018
Camera-ready deadline: June 21st, 2018
Workshop: July 13/14/15th, 2018
Reinforcement Learning (RL) agents traditionally rely on hand-designed
scalar rewards to learn how to act. The more complex and diverse
environments and tasks become, the more difficult it may be to engineer
rewards that elicit desired behavior. Designing rewards in multi-agent
settings with adversaries or co-operative allies can be even more
complicated. Experiment designers often have a goal in mind and then must
reverse engineer a reward function that will likely lead to it. This process
can be difficult, especially for non-experts, and is susceptible to reward
hacking---unexpected and undesired behavior that achieves high reward but
does not capture the essence of what the engineer was trying to achieve.
Moreover, hand-designed reward functions may be brittle, as slight changes
in the environment may yield large, and potentially unsafe, alterations in
The community has addressed these problems through many disparate approaches
including reward shaping, intrinsic rewards, hierarchical reinforcement
learning, curriculum learning, and transfer learning. Another approach is to
avoid designing scalar rewards altogether, and rather focus on designing
goals, for example, through inverse reinforcement learning, imitation
learning, target images, or multimodal channels such as speech and text.
This workshop will consider all topics related to designing goals for
reinforcement learning and problems that can arise from ill-defined goals.
The submissions can include novel research, open problems in the field, and
surveys. We are particularly interested in the topics of reward engineering,
reward hacking, interpretability, learning from humans and goal design using
AREAS OF INTEREST
Problems with reward design
- Robust reward functions
- Reward hacking
- Adversarial attacks on RL agents
- Generalizability of reward functions
- Communicating learned goals to humans
Submissions will be double-blind and are limited to 4 pages for short papers
and 8 pages for full papers, not including references and appendices.
Formatting should be in ICML style. Concurrent submissions are allowed, but
works that have been accepted at archival venues are discouraged.
Ashley Edwards, Georgia Institute of Technology
Himanshu Sahni, Georgia Institute of Technology
Kaushik Subramanian, Cogitai
Charles Isbell, Georgia Institute of Technology
Michael Littman, Brown University
Please address questions to: [hidden email]