IJCAI 2018 and SOCS 2018 summary
IJCAI 2018 and SoCS 2018 were held in Stockholm from July the 13th to July the 19th as a part of the Federated AI Meeting. Here I will give my personal summary comments on the conferences based on the talks I’ve seen. If anything is wrong or misrepresented, please don’t hesitate to get in touch with me.
I have not been to such a big scientific conference before, which was an interesting experience. In total all the conferences in FAIM had over 7500 participants, with around 2500 at IJCAI. SoCS on the other hand is a much smaller affair, with around 40 participants.
One thing that is quite difference for such a big conference is that there are very large companies having booths. The first days ICML (International Conference on Machine Learning) was still running, which meant that I got to see two sets of booths from different but related conferences. In general, it felt as if the ICML conference had more high-profile companies form a western perspective (Google and Deepmind, Netflix, Disney, …), while IJCAI had for me less known but still very large companies from Asia such as Tencent, JD, DiDi, and Alibaba Group. Not sure what to make of this difference.
In general I really liked the invited talks. In particular I would like to point out the following three talks
Yann LeCun gave a very interesting talk on the abilities and possibilities of deep learning and neural networks. The central idea is that we need to be able to learn not just a description of the world (perception) but a model of the world (reasoning).
Jean-Francois Bonnefon gave an interesting and thought-provoking talk on the moral machine, a viral web site designed to collect moral judgments from people around the world. A crucial moral problem for self driving cars and other AI-driven technologies is that they may need to decide whom to kill. In a car it may be the choice between the passengers, three jay walkers, or two law-abiding pedestrians.
Max Tegmark gave a wide ranging talk on AI and learning. Many interesting points, but one thing that stuck with me is his absolute refusal to be a part of creating any kind of Lethal Autonomous Weapons System, and his call to action to pledge to to be a part of such developments.
The Hallway Track
A common view for many conferences is that the so-called Hallway Track is often the best track. The Hallway track are the discussions with other people. As a lurker in the research community it is nice to meet old acquaintances and reconnect. I also met several new people and had very interesting discussions with them, including both learning new things and getting to give suggestions to others.
Sister Conferences Best Papers
One very nice thing is that IJCAI has several sessions with best papers from sister conferences such as ICML, ICLP, CP, and many more. This was very nice, and the talks I attended in these sessions were very interesting.
One really interesting paper was a machine learning paper that used the Arcade Learning Environment and learned evolution strategies instead. This gives much more interesting control structures in much less space.
Since it is my main research area, the constraint programming papers were of great interest to me. The most interesting one for me was the Compact MDD paper, which had several interesting contributions. I think it would be very interesting to see a combination of the ideas in this paper with the ones in Making Compact-Table Compact by Linnea Ingmar and my old advisor Christian Schulte.
I have a new found interest in game playing AI. The first day workshop was very interesting in general. In particular, I was quite inspired by the journal track paper MCTS-Minimax Hybrids with State Evaluations which extends MCTS with small shallow minimax searches to improve MCTS in games with strategic traps.
I’ve not thought much about path-finding before, except the standard Dijkstra for single-source shortest path and Floyd-Warshall all pairs shortest path. The Forward Search in Contraction Hierarchies paper by David Harabor and Peter Stuckey was very interesting and enlightening. The main thing that I had not thought about was that for repeated queries in a particular graph, pre-processing can be very important. Single queries are then 1 microsecond instead of 1 second. Contraction hierarchies is one such very interesting variant. Another interesting approach is |sub-goal graphs](/blog/post/conferences/faim-2018-7/#subgoal).
For path planning, the FastMap algorithm was quite interesting. It allows embedding a graph with edge-weights in Euclidean space, while maintaining similarity between node paths in the graph and node distances in the plane.
Something else that I had not thought about before is the multi agent path planning problem (MAPF). This problem considers a simultaneous path planning problem for many agents where collisions among the agents must be avoided. I will be looking into the constraint programming approaches that have been used for this problem.
Planning is an interesting domain, and it would be really fun to have the time to think about how to create the architecture for a planning system.