Over the past two weeks, I took some time out from work and visited two islands, one substantially larger in land area than the other (and, incidentally, I am returning to another island). I’ve been to both at some point in the past; the larger one for vacation, and the smaller for study, work, visiting family and friends and for vacation as well. The latter is not too difficult to guess – it’s Singapore; the former is Australia.
I was in Melbourne for the twenty-sixth International Joint Conference on Artificial Intelligence (IJCAI-17). Of course, the main highlight for me was presenting “Model Checking Multi-Agent Systems against LDLK Specifications“, a paper which built on parts of my Masters’ thesis on MCMAS-Dynamic. This was my first time presenting at an academic conference, though it’s not my first publication (that’s in AAMAS 17 on CTL*K); it was a little nerve-wracking, though I practiced the presentation a number of times beforehand and I think that helped a lot. Prof. Michael Wooldridge (the author of a commonly used textbook on multi-agent systems) even gave me a compliment that it was confidently presented. I’m not so sure about that, but I guess I have some private local state, and in any case deontic logic doesn’t typically have box p implies p as an axiom.
I also enjoyed the rest of the academic program as well. Apart from Alessio, I had the opportunity to meet up with several other academics whose names, at least, were very familiar. The keynote was delivered by Prof. Stuart Russell (one of the co-authors of the well-known Introduction to AI textbook), and discussed the concept of provably beneficial AI. Essentially, humans tend to be poor at specifying things (e.g. King Midas); many naive specifications of goals or objective functions naturally lead to agents enforcing self-preservation; more informally, “you can’t fetch the coffee if you’re dead”. The talk then discussed a way of formulating a problem such that it would always be a dominant strategy for an AI agent to allow itself to be switched off (as compared to not allowing itself to be switched off); see [Hadfield-Menell et al. 2016] (The Off-Switch Game) for more detail. The other invited speakers also had fairly interesting topics, discussing applications of AI in teaching, commerce and Texas hold-em poker.
I made it a point to attend as many paper presentation sessions as feasible. Expectedly, at a top conference like IJCAI there were quite a number of talks, especially in areas further removed from my work, that proved largely inscrutable. I think my presentation would also have been tricky to follow if one was not reasonably acquainted with temporal logic and/or multi-agent systems. Nonetheless, I gave my best effort in trying to follow the presentations, and there were certainly many interesting talks.
There was a session devoted rather specifically to the discussion of AI applied in the context of computer games. I found [Dann et al. 2017] (Real-Time Navigation in Classical Platform Games via Skill Reuse) to be particularly interesting; the key idea was having the agent acquire skills for small local movements, such as making a single jump to jump on top of a block three cells above the ground. The agent learns success probabilities for these skills, and uses a variant of Dijkstra with log-likelihood to find the “most likely” way to reach the goal safely (though there are some complications with re-planning if a step fails). The authors also demonstrated that the model performs better than the state of the art algorithm, which is based on A*, for maze-like levels. Other talks in the session also discussed level generation and physics engine inference (for platformers), RTS combat planning, narrative control in an educational RPG… and strategy derivation in Hex (this felt a bit out of place, but I guess they must have had exactly five papers that fit the theme well; sessions had either four or six presentations).
In terms of tutorials, I was unfortunately rather sleepy when I landed in Melbourne on Sunday. I did attend the tutorials on Monday, though; there was a very gently paced introduction to argumentation theory (the course Argumentation and Multi-Agent Systems which Prof. Francesca Toni teaches at Imperial would likely have covered this material, though she was on sabbatical when I was in year 4), and a less gently paced overview of heterogeneous learning (machine learning where some differences in the data sources can be exploited to yield better results).
Stepping back from the academic side of IJCAI, I also enjoyed the social events at the conference, and I think they were run fairly well; that said, I acknowledge I don’t have a benchmark at all. The reception at the Melbourne Cricket Ground was very generous, especially in terms of dessert (chocolate tarts!), and the conference in the Docklands was certainly done to a high standard. It also introduced me to the “alternating drop” protocol, which is apparently common (but peculiar to) Australia. Essentially, given two choices of main course X and Y, instead of having guests indicate their preference, they are served XYXYXY… around the table (of course, those with dietary restrictions are excluded). I can see how this is logistically advantageous (you know beforehand that it will suffice to prepare just over N/2 of both options, given N guests), but I’m not sure I like the probability of not getting my preference…
I stayed at the “recommended” Crown Metropol which was certainly luxurious. The hotel is located south of the Yarra river in Melbourne, and is part of the Crown Entertainment Complex – among other things, this features a fairly large casino, several restaurants and luxury shopping (way out of my budget range). Nonetheless, it was a good base from which I could easily explore the city, especially on the few days I stayed on after IJCAI.
I made it a point to leave a few days after the conference itself was over to explore the city, as I knew the conference days themselves were going to be fairly packed (8 am – 6.30 pm usually, running till 11 or so on the night of the Reception and Banquet). Like many work trips, the “official business” part of the trip was fairly draining – I perhaps should have seen this coming. I did manage to have a look around, though; I took a river cruise, walked through a few parts of the city and indulged in a bit of outlet shopping. I also sought out a local fast-food chain (Red Rooster, specialising in grilled chicken).
Given that I’d flown this far across the globe when going from London to Melbourne (it’s roughly 20 hours in the sky) I also decided to spend a few days in Singapore on the way back. For the most part I focused on spending time with family, though I also managed to find some time to flip through some of the IJCAI proceedings.
It’s surprisingly difficult to leave work behind…
This is certainly the first time I’ve had a two-week break since I started full-time at Palantir. (I had some things I decided I would clear out last year around the Christmas and New Year period, so no long break for me then.) That said, rather unsurprisingly in hindsight attending and presenting at IJCAI was easily just as intellectually stimulating and demanding as my regular work, so it felt more like a one-week break. Still, it was quite a dramatic slowdown.
I think the break worked well. It came at a time just after I launched some fairly hefty deliverables; knowing that I was leaving for two weeks, I had been pushing somewhat hard to get things out before I left. I hadn’t been back home for 11 months. I was initially a little apprehensive about going off for two consecutive weeks, though the logistics of course made sense, and I don’t think I have major regrets, at least not yet!