Schedule

Quasimetrics, invariance, and generalization in RL

Benjamin Eysenbach

Getting complex behavior from intrinsic motivation to occupy future action-state path space 

Jorge Ramírez-Ruiz

Most theories of behavior posit that agents tend to maximize some form of reward or utility. While the source of this reward can be external (e.g. training a robot to do something), or internal (e.g. homeostasis) to the agent, arguably a successful theory of behavior that is based on reward-maximization needs to integrate the evolutionary, ecological and developmental constraints of the agents to explain why rewards are rewards. In this talk, I propose an entropic principle framing behavior as occupying future paths of actions and states, which is meaningful only after considering the agent's constraints. According to this maximum occupancy principle, reward signals, even if internally generated, are the means to occupy path space, not the goal per se. Goal-directedness thus simply emerges as rational ways of searching for resources so that movement, understood amply, never ends.

Curiosity in the Natural and Artificial Wild

Dani Bassett

When we are curious, how do our minds move through the complex landscape of possible ideas? In this talk, I will first describe a historico-philosophical categorization of curiosity styles and provide evidence for those styles both in the laboratory and around the world, before then discussing efforts to design artificial agents who build knowledge networks using the same movement styles.

The Curious U: Integrating Theories Linking Knowledge and Information-Seeking Behaviour

Alexandr Ten

In this talk, I am going to share a bird’s-eye perspective on several theories that contain explanations about a non-monotonic (inverted-U) relationship between knowledge and curiosity. From this perspective, the inverted-U relationship can arise as a consequence of knowledge optimization under specific constraints. I will also discuss some existing simulation and human-behavior studies in light of this perspective.

Lightning talks, Round 1

Moderator: Surabhi S Nath

Lightning talks, Round 2

Moderator: Surabhi S Nath

Using games to study intrinsic motivation

Franziska Brändle

What makes an activity intrinsically rewarding? We propose a novel approach to answer this question: Using video games. In this talk, I will present a project focused on learning progress, to demonstrate the potential of video games as tools for studying intrinsic motivation in ecological and engaging settings.

The nature of intrinsic motivation

Doina Precup

Panel: Unifying Principles of Intrinsic Motivations

Eysenbach, Bassett, Ramírez-Ruiz, Precup, Brändle

Moderator: Alireza Modirshanechi