Projects

Below are short summaries of my work roughly organized into broad projects. For a complete list of papers, see here.

Miming involves communicating with your actions

Motivation

Observing someone's actions allows us to infer what they know, expect, want, or other mental states. Conversely, the person being observed could use their actions to show an observer what's in their head. This can occur during everyday teaching or in other social interactions and activities (e.g. miming).

Summary

We examine communicative demonstrations using a combination of computational models and experiments. In particular, we study how people optimally plan over observer belief transitions to accomplish communicative goals as well as how recognition of a demonstrator's communicative goals affects interpretion by an observer.

Relevant Work

Code

Two (non-human) agents receiving rewards/punishments

Motivation

Social rewards and punishments are everywhere. They can occur between a parent and children or in other social interactions (even those involving non-humans), and they have been proposed as a way to teach artificial learners. But how exactly do people use rewards and punishments to influence another's behavior?

Summary

It is often assumed that people use rewards and punishments to shape others' behavior (e.g. in the sense of operant conditioning). However, through a combination of computational and emprical studies, our work argues that people do not reward and punish in this manner. Rather, rewards and punishments tend to be used communicatively to signal information. This insight motivates a new perspective on human social learning as well as new interactive machine learning algorithms.

Relevant Work

Code

Motivation

When we observe someone else's behavior, we can often infer how they represent a task. This requires the observer to have a rich theory of an actor's unobservable mental structure.

Summary

Using ideas from computer science and reinforcement learning, we examine how people and computers can interpret others' behavior in terms of rich representational structure.

Relevant Work

Motivation

Norms are often represented abstractly rather than in terms of low level actions. This raises the question of how such abstract representations can be jointly learned through social interaction.

Summary

Using the framework of multi-agent reinforcement learning and multi-participant experimental paradigms, we explore how people learn and use abstract norms. In particular, we examine norms of cooperation/competition as well as behavioral norms that can arise in fully collaborative settings where people must work together to achieve a shared goal.

Relevant Papers

Code

Last updated: Mon 01 October 2018

Copyright © Mark Ho 2016-2018. Theme based on Pelican-Hyde.