Guest post from Ariel Procaccia:
IJCAI is held biennially in odd-numbered years since 1969, so IJCAI’09 is the 21st in the series (AAAI takes place every year, with the frequent exception that there is no AAAI when IJCAI is in North America, e.g., this year). Generally speaking, AAAI/IJCAI are considered the top conferences in AI; in case you were wondering how one defines AI, my best definition is “everything that might get published in AAAI/IJCAI”. For a more detailed (and less cyclic) definition see the list of keywords in the call for papers. IJCAI’09 had 1290 submitted papers and 331 accepted papers.
The conference included plenty of Algorithmic Game Theory papers. I consider at least 7 sessions (with 4 papers per session) to be “hardcore” Econ/CS sessions: three social choice sessions (!!), coalitional games, solution concepts in games, auctions, and mechanism design. Some AGT papers appeared in indirectly related sessions about social networks, “performance and modeling in games”, etc.
A short comparison to the previous week’s EC is called for. One point is that social choice and coalitional games are very popular in IJCAI. In contrast, it seems that the vast majority of papers in EC were concerned with mechanism design, often in the context of approximation. Furthermore, the emphasis in EC seemed to be on technical papers, whereas the main emphasis in IJCAI (and AI conferences in general) is on the conceptual aspects rather than the technical, which usually makes the conference fun to attend. That said, I personally think that the technical threshold is satisfactory, and one often encounters mathematically involved papers (such as a paper by Lirong Xia and Jérôme Lang with a 15 page proof). Finally, EC had only 7 accepted papers labeled as “AI”, and in my opinion only one or two of them were “real” AI (for example, my own EC papers were labeled as “AI” as I consider myself to be an “AI guy”, but were not “real” AI). For more see Shahar’s EC conference report of July 11, which correctly observed that “AI guys complain that there are too many theory papers”, although I don’t see what made the theory guys complain.
Turning back to IJCAI, the keynote address was given by Hal Varian, who spoke about computer mediated transactions. Varian did an interesting job of putting the recent Internet-related developments in perspective by providing many historical anecdotes. I especially liked his discussion of how better monitoring is related to better contracts. For instance, as early as 3000BC clay tokens were matched to jars in order to guarantee honest delivery between Mediterranean countries, thereby significantly improving trade. In his introduction Tuomas Sandholm noted that HAL is an appropriate name for a speaker in an AI conference; Varian replied that in a recent statistics conference someone said that he was invited since his name is Hal Variance.
Let me briefly mention one game theory talk that I particularly enjoyed. A paper called “Iterated regret minimization: a new solution concept”, by Joe Halpern and Rafael Pass (Cornell), was nicely presented by Rafael. The paper defines iterated regret minimization in the natural way (given that you know what regret minimization is). Most interestingly, it turns out that in several prominent games where standard solution concepts (Nash, iterated elimination of dominated strategies, etc.) give absurd outcomes, iterated regret minimization provides intuitive outcomes that agree with empirical experiments in a very satisfying way. Joe and Rafael also provide existence theorems and an epistemic characterization.
I will conclude with some assorted highlights. First, the important-sounding “AAAI Presidential Panel on Long-Term AI Futures” was charged with dealing with the seemingly sci-fi issues that might arise when we reach the “singularity”, that is, the point where machines are as intelligent as humans. The presentation was somewhat—how should I put it?—bizarre unexpected, but nevertheless thought provoking. They actually discussed (among many other issues) Asimov’s laws of robotics. Here is some food for thought: how do you specify and implement Asimov’s first law (a robot cannot harm a human)? If a robot makes a child cry by trying to help the parents, does it violate this law? It turns out that there are some AAAI/IJCAI papers on this issue! Second, the titles of the outstanding papers are “Learning conditional preference networks with queries” and “Consequence-Driven Reasoning for Horn SHIQ Ontologies”; the readers of this blog will surely excuse me for not elaborating further. Third, it seems the zeitgeist is AI and the environment, not surprisingly given Obama’s funding policies; environmental game theory is probably the next big thing. Fourth, check out this movie of how Andrew Ng of Stanford University (co-winner of this year’s IJCAI Computers and Thought award) used machine learning to build an awesome autonomous helicopter.
The next IJCAI will take place in Barcelona in 2011. It was also announced that IJCAI’13 will be held in Beijing. By then China will take over the universe, so perhaps we will not need a visa.
Now I understand why machine-learning has split up from AI and the sensible people have moved to machine-learning.
AA, which part of the post made you reach this amazing off topic enlightenment?
I admit the highlights part (panel, environment) was a bit sarcastic, but this is exactly the source of IJCAI’s charm: the conference is truly interdisciplinary and thought provoking.
I also disagree with the statement “machine learning has split from AI”. It is true that there are dedicated machine learning conferences such as NIPS and ICML, but this is also true for other fields that are “hardcore” AI such as multi-agent systems (AAMAS) and planning (ICAPS). Indeed, some of the top machine learning people regularly publish in AAAI/IJCAI; see for example the DBLP of Andrew Ng who was mentioned in the post (47 NIPS/ICML papers and 15 AAAI/IJCAI/UAI papers).
[…] were arguably those of Robotics and Learning. Whereas all have a significant overlap with AI (see Ariel’s post on IJCAI for a definition), these three are considered to be almost mutually exclusive in terms of their […]
[…] were arguably those of Robotics and Learning. Whereas all have a significant overlap with AI (see Ariel’s post on IJCAI for a definition), these three are considered to be almost mutually exclusive in terms of their […]
[…] a working definition of AI, “AI is whatever gets published at AAAI/IJCAI” similarly to Ariel Procaccia stating “in case you were wondering how one defines AI, my best definition is everything that might get […]