One of the key problems in reinforcement learning is exploration. One of the ways to solve it is to introduce individual curiosity to the agent. However, authors of a recent paper follow a different logic: they try to imitate exploration done by human tribes. This exploration was driven by external motivation - changing weather conditions.
Paper introduces a multi-agent approach with populations dynamic. Every population is guided by its own policy. The reward is averaged across the population as well.
Authors show improved performance in tasks that require specialization from populations.
Speaker: Mikita Sazanovich.
Presentation language: Russian.
Date and Time: February 19th, 18:30-20:00.
Place: Times, room 204.
Videos from previous seminars are available at http://bit.ly/MLJBSeminars