IBM Research Dublin, winner of the Shifts Challenge 2023: Interview

  • Insights

As the shipping industry grapples with the challenges of rapid digitalisation, it’s more important than ever that we strive for ever greater accuracy, robustness and transparency when it comes to new technology.

This year it was an expert team from IBM Research in Dublin that won the Shifts Challenge 2023, presenting a strong approach to modelling vessel behaviour using AI.

In this interview with Shivani Tomar, Pre-Doc AI Researcher at IBM Research, we delve into the team’s journey, from initial formation to finding inspiration in the wind turbine sector – and how this sort of research is vital to making the shipping industry more efficient. Join us as we discuss the facts behind their winning solution, and explore the real-world implications of their work.

Q: Tell us a bit about your team and how you came together to tackle the problem of distributional shift in the shipping industry as part of the Shifts Challenge 2023.

At IBM Research Dublin, we have a team of leading industry experts continuously working across a number of different research areas such as time series forecasting, incremental machine learning and privacy enhancing technologies. Essential to the ongoing success of the team is our culture of collaboration and continuous sharing of latest insights and information. It was this team approach that proved critical to developing a winning solution.

Q: Your novel approach applied learnings from another sector unrelated to shipping. Can you explain how the wind turbine sector’s learnings were applied to enhance vessel power estimation?

There was a stark similarity in the data relating to the wind turbine sector. We observed similar trends in the wind features as the wind-based components in the power estimation data shared by the challenge organisers. Since the distributional shift was introduced by partitioning the data across wind components, it led to the obvious choice of training specialised ensembles catering to each wind partition. The wind models that we developed were partially funded by the EU Horizon 2020 program through the MORE (Management of Realtime Energy data) project.

Q: Assessing the quality of an AI model is a major problem and critical to getting people to trust AI. DeepSea deployed a novel evaluation benchmark to help solve this challenge as part of the Shifts Challenge. What did you think about DeepSea’s methodology and approach?

The evaluation benchmark provided was quite useful, however, it only considered the extreme data distributions for evaluation purposes. We believe multiple scenarios and splits would probably be more useful to test rather than one scenario of unseen evaluation data.

Q: The reference model provided for the challenge set a benchmark. Your team achieved a 13% improvement on vessel power against the reference model provided. What were the key findings and performance improvements that you observed?

A part of the improvement came from the obvious starting points when trying to improve an existing benchmark like hyper-parameter optimisation. However, a major improvement in the results can be attributed to the cross-application of the partition-based approach from the wind turbine sector.

Q: Beyond the quantitative improvements, what are some potential real-world implications of your work? How can your approach contribute to the shipping industry’s efficiency, sustainability, or cost-effectiveness?

Our results are a step closer towards the integration of AI solutions to real world applications i.e., the shipping industry in this case. Based on the improvements achieved with our approach, we can, through collaboration, bring down the costs associated with inaccurate power prediction like errors in fuel planning and route optimisation.

Shivani is a PhD Student working in collaboration with IBM Research and Trinity College Dublin. Her research area includes incremental machine learning mainly focussing on time series data. Shivani’s current interests involve explainability in time series regression problems using prototypes in streaming context.