Summary

The big news since the last update is that I’ve started full-time employment as a Modelling and Simulation Engineer at Inovor Technologies. Inovor is an Adelaide-based company that designs and builds CubeSats for customers such as CSIRO. I am involved in a number of exciting projects, including aspects of the attitude determination and control system (ADCS). I’m really enjoying this opportunity to contribute and learn.

My study topics discussed on this website may now include some theory related to attitude determination and control, but I won’t be discussing specifics to be sure I don’t breach any confidentiality agreements. This website will continue to focus primarily on machine learning projects.

Full-time work comes with reduced hours for other projects. My involvement in the AI Collaborative Network Data Collaboration and the Adelaide University Solar Race Team has now dropped off. I would like to contribute further to each of these projects, but I’m not sure how much I will manage.

My change of circumstances makes it more important, again, to prioritise any outside projects and tasks. The technological project I am most interested in continuing to pursue outside of work is the development of constructive algorithms for neural networks. The exact approach I take to this is still in flux.

The project of developing tutorials and a library that people can use for machine learning interests me, but this should come after I have achieved more success with implementing, creating and applying these algorithms to machine learning problems. I am yet to fully resolve whether I continue to work with PyTorch or develop my own code for the neural network.

Applications in control and robotics are a of particular interest to me. These are often cast as reinforcement learning problems, but can incorporate supervised and unsupervised learning. I will try to build constructive algorithms compatible for tasks in each type of learning.

I had plans to write papers for the Australasian Conference on Robotics and Automation held in Adelaide later this year, but with my new employment this is looking less likely. Nevertheless, I will keep the submission date in mind for some extra motivation to get some work on autonomous learning and control to a conference paper ready state.

I will try to continue writing about other topics on my Wordpress site (https://tobylightheart.wordpress.com).

Updated Project and Task List

Neural Networks: Construction and Selection

  • Review current progress and development plan
  • Literature and Study
    • Constructive neural networks
    • Meta-learning

Research Conferences/Competitions

  • Reinforcement learning with neural network construction for visual control
    • Initial literature survey
    • Theory development plan
    • Code development plan
    • Submission deadline: 2019-09-23
  • Self-supervised motion control learning with reinforcement-based target selection
    • Initial literature survey
    • Theory development plan
    • Code development plan
    • Submission deadline: 2019-09-23
  • NeurIPS competitions (https://nips.cc/Conferences/2019/CallForCompetitions)
    • Update and test Malmo (https://www.aicrowd.com/challenges/neurips-2019-minerl-competition)
    • Register for MineRL Competition
    • MineRL first round submissions close: 2019-09-22

Other Projects

  • Adelaide University Solar Race Team
    • Develop GUI for telemetry and predictions.

Selected Tasks

  1. Review current progress and development plan
  2. Survey recent literature on constructive neural networks and reinforcement learning for physics simulations/robotics
    1. Conference paper: Reinforcement learning with neural network construction for visual control
    2. Conference paper: Self-supervised motion control learning with reinforcement-based target selection
  3. Update and test Malmo