Publications

Journal Papers:

  • Where’s the Reward? A Review of Reinforcement Learning for Instructional Sequencing
    Shayan Doroudi, Vincent Aleven, Emma Brunskill
    International Journal of Artificial Intelligence in Education
    Publisher’s Version
    Author’s Version

Conference Papers:

  • Robust Evaluation Matrix: Towards a More Principled Offline Exploration of Instructional Policies
    Shayan Doroudi, Vincent Aleven, Emma Brunskill
    In Proceedings of Learning @ Scale (L@S) 2017.


    It would be useful if we could tell how well a new teaching strategy would do before running an experiment. It would be even better if we could determine what is the best instructional policy to implement without having to test a bunch on actual students. One way to do this is to simulate a new instructional policy using a hypothesized student model. Unfortunately, this could lead to overestimating the performance of new teaching methods. We propose an alternative approach: using many models we expect to be wrong, rather than using one model we hope to be right.

  • Sequence Matters, But How Exactly? A Method for Evaluating Activity Sequences from Data
    Shayan Doroudi, Kenneth Holstein, Vincent Aleven, Emma Brunskill
    In Proceedings of Educational Data Mining (EDM) 2016.

    The order in which material is presented can strongly influence what is learned, how fast performance increases, and sometimes even whether the material is learned at all…as we discover the underlying principles of the order effects in learning, we move instruction away from idiosyncratic expression and closer to a controlled and predictable science.
    Frank E. Ritter and Josef Nerb, In Order to Learn

    We’re still not there yet! I think this paper’s a step in that direction!

  • Toward a Learning Science for Complex Crowdsourcing Tasks
    Shayan Doroudi, Ece Kamar, Emma Brunskill, Eric Horvitz
    In Proceedings of Computer Human Interaction (CHI) 2016.
    (Supplementary Materials)


    Can we teach others how to solve a task effectively without knowing how to solve it ourselves? I think so! Our work shows that we can effectively train crowd workers in solving complex web search tasks by having them review their peers’ solutions. If we only present workers with solutions that are beyond some threshold length, reviewing peer solutions can surpass the efficacy of expert examples. This work is one step towards what may hopefully become a new era of crowdwork, where workers can be trained to solve complex tasks that cannot easily be decomposed. It also suggests a way to form low-cost yet effective educational materials for other educational settings. It turns out a lot of workers find complex web search tasks pretty fun too!

    I have done over 18,000 HITs on mTurk, and your HIT was, by far, the most interesting, perfectly challenging, and FUN HIT I have ever done on mTurk. While I appreciate the bonus for the HIT, what I appreciate more was the actual HIT itself. Thank you. Thank you so much for bringing it to mTurk. I clicked the link to the daily Google a day thing, and am going to do it every day. I loved it that much. 

  • A PAC RL Algorithm for Episodic POMDPs
    Zhaohan (Daniel) Guo, Shayan Doroudi, Emma Brunskill
    In Proceedings of Artificial Intelligence and Statistics (AISTATS) 2016.

Workshop Papers:

You can also follow me on academia.edu.