See Archive

Deep Reinforcement Learning in High-Dimensional Linear Action Spaces / Socially Responsible Design Science Research in Information Systems

Title: Deep Reinforcement Learning in High-Dimensional Linear Action Spaces – talk given by Wouter van Heeswijk

  • Abstract: Recent years have seen a surge in popularity in the field of Reinforcement Learning (also known as Approximate Dynamic Programming in the Operations Research domain), not in the least due to great advances made in the area of neural networks. In reinforcement learning, we seek to learn some unknown function that tells us what the expected future impact is of an action performed today. As neural networks may potentially capture every continuous function, the reason for its appeal in Reinforcement Learning is eminent.

    However, to select the best possible action out of a set, the output of the neural network should be recalculated for each possible action. As Operations Research typically deals with high-dimensional action spaces containing dazzling numbers of combinations, full enumeration is computationally intractable even for modest problem instances. Linear programming has proven very successful in dealing with such problems, yet scales poorly to multi-period decisions that embed dynamic and stochastic properties, as is often the case in real-life problems.

    This talk addresses the integration of linear programming and neural networks - which by definition perform nonlinear transformations - in the context of Reinforcement Learning, aiming to solve problems with high-dimensional action spaces. The hybrid solution method is illustrated by means of a simple yet computationally challenging container shipping problem. I hope to provide some insights into the challenges of dimensionality and how combinations of existing techniques may help to overcome these challenges.
  • Bio: Wouter van Heeswijk is an assistant professor within the IEBIS department, active on the intersection of Financial Engineering and Operations Research. He received his PhD within the same department in 2017 under the supervision of Prof. Henk Zijm. Aside from his affiliation with the University of Twente, he has been a visiting researcher at the Technical University of Denmark (DTU) and a postdoctoral researcher at Centrum Wiskunde & Informatica. He performed research on a variety of topics including transportation management, real option analysis, combinatorial optimization, multi-agent simulation and reinforcement learning, currently focusing primarily on the latter.

Title: Socially Responsible Design Science Research in Information Systems – talk given by Mike Monson

  • Abstract: “The practice of critical IS research is characterised as a wide range of diverse research endeavours aimed at revealing, criticizing and explaining technological developments and the use of IS in organisations and society that, in the name of efficiency, rationalisation and progress, increase control, domination, and oppression, and produce socially detrimental consequences” (Cecez-Kecmanovic 2007, p1446).

    So what?

    Should the IS discipline be concerned?

    If so, where do we fit into the ‘bigger scheme of things’?

    The proposed solution:
    Engage in research activities that limit the negative and provide positive impact.

    My research aims to produce prescriptive studies to design and develop methodological frameworks for the conduct of socially responsible design science in the IS field.
  • Bio: I have worked in IS for almost 5 decades. I was an entrepreneur and business owner in the software development industry. I was educated at the University of Cape Town with degrees in Finance and IS. My current focus is research methodologies for socially responsible design science research.