purplecat

I have a publication in the Agents and Robots for Reliable Engineered Autonomy (AREA) workshop. Corroborative V&V for Autonomous Systems: Integrating Evidence and Discrepancy Analysis for Safety Assurance. It should be open access, but does not appear to be. It's not a super-exciting paper. It takes the observation that, if you are doing assurance of robotic systems you will take a variety of approaches; abstract models, simulated tests, hardware tests... and then have to reconcile the results of these approaches. The paper describes the first stab at a tool for this, but it is a very early prototype.

My PhD student had a paper published in AAMAS on Uncertain Machine Ethics Planning. This is a good conference which, for my sins, I'm currently joint Programme Chair for (this means I'm currently in the process of trying to find 1,300 potential referees in the hopes of ending up with 650). Anyhoo... AAMAS rewards pretty theory heavy papers and this was no exception, but the bottom line is that he's developed a technique in which a system can reason across several potential plans of action, using different moral theories in order to work out which plan of action is least unacceptable across all the moral theories (I hope this makes sense, we keep running into double negatives in the theory). It's grounded in a philosophical concept called hypothetical retrospection - in which even if something turns out badly you can argue it was still the correct choice because at the time you made the choice the chance of it turning out badly was low. There are some details such as ranking outcomes so, in the situation where you can get an apple (for sure) or gamble with a low chance on getting an all expenses paid holiday (yes I know this isn't a moral choice), no number of apples can outweigh the small chance of getting the holiday - I guess the moral equivalent might be no number of people made a little bit happier can be outweighed by killing someone.

Moral theories can be big theoretical juggernauts like utilitarianism or kantian morality - or more subtle distinction around which values are preferred (though this doesn't really come out in the paper if you can wade through all the formalism).

The award winning paper I mentioned next week, actually had a sequel. In The human factor: Addressing computing risks for critical national infrastructure towards 2040 we performed a similar exercise of asking a number of experts about risks to Critical National Infrastructure arising from computing developments and synthesising the results.

I am honestly, happier with this paper, I thought we had a better range of genuine expertise in the people we talked to, and a more focused area of consideration. We had a little trouble with the third referee, who thought our experts were wrong about Quantum Computing and that we should rewrite the paper so they gave the answer the referee thought was correct. Our experts did not think Quantum Computing was among the biggest risks to be considered in the next 15 years - but instead thought there were a number of issues relating to human factors (sophisticated phishing, difficulty tracing the cause of problems and poor incident response in complex situations).

The work I'm involved with on the CRADLE Project involves trying to put together an assurance case for an automonous robot to be deployed... somewhere. At the moment the various bodies like HSE, The Office for Nuclear Regulation, The Civil Aviation Authority and so on, don't really want to be drawn on what evidence they would need to be certain due diligence had been performed for the deployment of an autonomous robot. We're therefore trying to produce some evidence that a robot is safe and see if they might at least admit they wouldn't throw such evidence out immediately.

We're also interested in how the production of such evidence could be made more stream-lined to avoid having to come up with new processes each time a different robot is considered. Hence Towards Patterns for a Reference Assurance Case for Autonomous Inspection Robots (which doesn't appear to be Open Access even though it should be, but I'm pretty sure anyone reading this here knows how to contact one of the authors and request a copy, should they be interested)

Here are some photos of my new lab at work:

( Photos under the Cut )

My colleagues in Engineering who three years ago were all moved into a ginormous custom-built state of the art building where there is not enough room and they have to share offices are all a bit gob-smacked by how I managed to get this much space. I'm not honestly sure I know either - I forsee spending the rest of my career defending my territory.

The department recently partnered with the British Computer Society (BCS) to make some videos for... I don't know, honestly, generic publicity reasons I presume. Anyway the final results prominently feature one of my postdocs (Hazel) and my former boss (Michael). I get a bit part in Hazel's video and the book Michael and I wrote features in his video.

The interested can find them here.

I was going over my early LJ posts and regretting that I don't post about science much any more. Science blogging requires a fair bit of mental energy to put everything together in a coherent, yet understandable, way.

Anyway, here is a paper in which I, another Computer Scientist and a legal scholar discuss what kind of explanations offered by autonomous and AI systems might be of interest to lawyers. This includes both explanations that might be offered to lawyers after some event, but also what lawyers might want to know about explanations offered to users during some interaction that let to an event. It's fairly preliminary work aimed at mapping the space more than offering specific conclusions or calls to action.

Verifiable Autonomous Systems

I wouldn't expect anyone here to buy it, and certainly not at the price CUP are charging. But still, BOOK!

Model-Checking for Analysis of Information Leakage in Social Networks has been accepted into the post-proceedings volume of the conference it was given at. The acceptance is notable, in particular, for the referees' comments which amount in all cases to "have dealt with previous comments more than sufficiently, nothing more to do" which is good in the light of how much time I don't have between now and the camera ready deadline.