IN-PERSON EVENT - brum.ai May 22 - Explainable RL & Constraint-Based Programming
Details
May 2022 - in-person @Huel Birmingham
Software Engineering for Explainable Reinforcement Learning (XRL)
Constraint-Based Programming
Also, online version (beta) parallel event
👉 https://www.meetup.com/brum-ai/events/285970650/
COVID-19 safety
Optional, wear a mask! (I do)
Stickers with different colors to signal your social-distancing preferences.
Capacity is restricted preventing overcrowding.
Sanitser will be provided.
Location
We are very excited to be back in person and hosted by Huel UK in their new Birmingham base in The Custard Factory https://careers.huel.com/locations/huel-birmingham. Meet at the Gibb Street entrance, https://w3w.co/flips.opera.even look for the Huel Tshirt!
Capacity
In-person spaces are restricted, please only book if you can attend!
It will feel really good to have a full room so please don't reserve a seat unless you are coming, the event is in-person.
Please update your RSVP if something changes, you may be preventing someone else from attending if there are no-shows.
We have agreed the capacity with the host.
Agenda
* New start time 18:15
Please arrive 18:15 - 45 as we can't guarantee we could let people in after 18:45 due to door access
18:15 Pizza and drinks
18:45 - 19:20 Talk 1 + Q&A
19:20 - 19:25 Speaker Intermission
19:25 - 20:00 Talk 2 + Q&A
Drinks at a local to follow.
Event overview
T1: The adoption of AI and Machine Learning (ML) has become ubiquitous in modern software. Reinforcement Learning (RL) is a particular sub-field of ML with great success in applications such as self-driving cars, industrial automation, and finance systems, among many others. Despite its broad applicability, the nature of RL is still considered a “black-box” where system decisions can become opaque to stakeholders. The insufficiency of validation techniques for the reasoning done by the system when using RL is a deterrent to broader adoption. Explaining the decision-making processes becomes increasingly important to enhance collaboration, and to increment confidence in these types of systems. This is ratified by the General Data Protection Regulation (GDPR) law, which enshrines the right to explanation.
Juan will present ongoing research on applying the SE approaches of Model-Driven Engineering (MDE) and Event-Driven Monitoring (EDM) for explainability in RL.
T2: Constraint-based programming takes a heuristic approach to solving a variety of problems, and whilst constraint-based programs are typically going to be sub-optimal, often enough, a “satisfactory” solution is all that’s needed. Within this domain, hundreds of hours can be devoted to the creation of a bespoke and highly tuned algorithm, one that is able to exploit the details of a problem to improve a yield or increase efficiency. But it is this specificity which will make the algorithm inevitably hard to evolve, and much more worryingly, hard to understand. Fortunately, we can expediate the tuning of these algorithms, deferring the modelling of these problems to highly optimised solvers that can do 80% of the hard work for 2% of the cost.
But...
- Why should people care about constraints?
- How do these solvers work?
- Are they the “be all, and end all”?
The answers to these questions will be followed by a discussion on the implementation of MILPS (constraint programs) to model trained neural networks, a method which enables guarantees about the optimality of an algorithm.
Speaker Bios
Juan Marcelo’s research career started in his home country Ecuador in 2015 where he applied Internet of Things (IoT) and SE techniques to build ambient assisted living solutions for the elderly. In 2018, he came to the UK to start his PhD at Aston University researching SE techniques for enabling explainability in autonomous self-adaptive systems (SAS). Currently, he works as a senior research associate at the Smart Internet Lab at the University of Bristol.
- https://scholar.google.com/citations?user=eFeGkgQAAAAJ&hl=en&oi=ao
- https://www.linkedin.com/in/juan-marcelo-parra-ullauri-7b27a313a/
Daniel Allford, of circuitmind.io, is a lead software engineer with years of experience spanning a variety of roles that consistently utilise cutting-edge technologies. From data science, data migration and automation, Daniel’s primary interest is now within constraint-based programming, a paradigm dedicated to solving combinatorial problems.
https://www.linkedin.com/in/daniel-allford-901b233a/.
[https://medium.com/@d-allford](https://www.linkedin.com/safety/go?url=https%3A%2F%2Fmedium.com%2F%40d-allford&trk=flagship-messaging-web&messageThreadUrn=urn%3Ali%3AmessagingThread%3A2-MDkxMmE0ZWMtMDk1Zi00NTBjLTk5OWQtZmE1ZWY0NTdhNzZjXzAxMg%3D%3D)