Reinforcement Learning: Efficient Communication and Sample Efficient Learning

Doktorsavhandling, 2024

Life is full of decision-making problems where only partial information is available to the decision-maker and where the outcomes are uncertain. Whether it’s choosing a restaurant for dinner, selecting a movie on a streaming service, or conveying concepts during a lecture, the decision-maker observes only the results of their choices without knowing what would have happened if it had acted differently. Because of this, the decision-maker needs to carefully balance between using its current knowledge, to make good decisions, and exploring the unknown to gather new information that might lead to even better decisions in the future.


In this thesis, we explore several topics in reinforcement learning - a computational approach to sequential decision-making under uncertainty. The first part investigates how efficient communication emerges between reinforcement learning agents in signaling games. The support for efficient communication, in an information-theoretic sense, is an important characteristic of human languages. Our agents create artificial languages that are as efficient as human languages as well as similar to human ones. We also combine reinforcement learning with iterated learning and find that this combination accounts better for human color naming systems than what any of the models do individually.


The second part focuses on sample-efficient algorithms for multi-armed bandits. We propose Thompson sampling-based methods for regret minimization in multi-armed bandits with clustered arms. Additionally, we address finding optimal policies with fixed confidence in bandits with linear constraints. For this problem, we characterize a lower bound and illustrate how it depends on a non-convex projection onto the normal cone spanned by the constraints. We leverage these insights to derive asymptotically optimal algorithms for pure exploration in bandits with linear constraints. Finally, we apply techniques from multi-armed bandits to develop active learning strategies for ordering items based on noisy preference feedback.