Benjamin Rosman

My research interests focus on aspects of Intelligent Robotics, particularly intelligent decision making and learning deep structure in strategies for surviving in nonstationary and adversarial worlds. This includes investigating approaches to automated discovery of novel concepts, through state and action abstraction, as well as representation learning. A major theme of my work is transfer learning.

See my Google Scholar profile.

---

I am interested in lifelong learning agents, and in particular the idea of learning "common sense" knowledge by abstracting behavioural commonalities from a number of different tasks in the same domain.

1. Action Priors for Learning Domain Invariances
2. Learning Domain Abstractions for Long Lived Robots [thesis]
3. Behavioural Domain Knowledge Transfer for Autonomous Agents
4. What Good are Actions? Accelerating Learning using Learned Action Priors
5. A Multitask Representation using Reusable Local Policy Templates
6. Giving Advice to Agents with Hidden Goals

---

Another main focus of my work is on the theme of online personalisation and adaptation of interfaces and content.

1. Learning Domain Abstractions for Long Lived Robots [thesis]
2. Adapting interaction environments to diverse users through online action set selection
3. Latent-variable MDP models for adapting the interaction environment of diverse users
4. On user behaviour adaptation under interface change

---

I am also interested in Computer Science education, and more generally the teaching of scientific concepts, particularly in the context of South Africa.

1. Language Performance at High School and Success in First Year Computer Science