My response to Ben's paper is to be cautious about drawing conclusions from simulated environments. Human level AGI has an algorithmic complexity of 10^9 bits (as estimated by Landauer). It is not possible to learn this much information from an environment that is less complex. If a baby AI did perform well in a simplified simulation of the world, it would not imply that the same system would work in the real world. It would be like training a language model on a simple, artificial language and then concluding that the system could be scaled up to learn English.
This is a lesson from my dissertation work in network intrusion anomaly detection. This was a machine learning task in which the system was trained on attack-free network traffic, and then identified anything out of the ordinary as malicious. For development and testing, we used the 1999 MIT-DARPA Lincoln Labs data set consisting of 5 weeks of synthetic network traffic with hundreds of labeled attacks. The test set developers took great care to make the data as realistic as possible. They collected statistics from real networks, built an isolated network of 4 real computers running different operating systems, and thousands of simulated computers that generated HTTP requests to public websites and mailing lists, and generated synthetic email using English word bigram frequencies, and other kinds of traffic. In my work I discovered a simple algorithm that beat the best intrusion detection systems available at the time. I parsed network packets into individual 1-4 byte fields, recorded all the values that ever occurred at least once in training, and flagged any new value in the test data as suspicious, with a score inversely proportional to the size of the set of values observed in training and proportional to the time since the previous anomaly. Not surprisingly, the simple algorithm failed on real network traffic. There were too many false alarms for it to be even remotely useful. The reason it worked on the synthetic traffic was that it was algorithmically simple compared to real traffic. For example, one of the most effective tests was the TTL value, a counter that decrements with each IP routing hop, intended to prevent routing loops. It turned out that most of the attacks were simulated from a machine that was one hop further away than the machines simulating normal traffic. A problem like that could have been fixed, but there were a dozen others that I found, and probably many that I didn't find. It's not that the test set developers weren't careful. They spent probably $1 million developing it (several people over 2 years). It's that you can't simulate the high complexity of thousands of computers and human users with anything less than that. Simple problems have simple solutions, but that's not AGI. -- Matt Mahoney, matmaho...@yahoo.com --- On Fri, 1/9/09, Ben Goertzel <b...@goertzel.org> wrote: > From: Ben Goertzel <b...@goertzel.org> > Subject: [agi] What Must a World Be That a Humanlike Intelligence May Develop > In It? > To: agi@v2.listbox.com > Date: Friday, January 9, 2009, 5:58 PM > Hi all, > > I intend to submit the following paper to JAGI shortly, but > I figured > I'd run it past you folks on this list first, and > incorporate any > useful feedback into the draft I submit > > This is an attempt to articulate a virtual world > infrastructure that > will be adequate for the development of human-level AGI > > http://www.goertzel.org/papers/BlocksNBeadsWorld.pdf > > Most of the paper is taken up by conceptual and > requirements issues, > but at the end specific world-design proposals are made. > > This complements my earlier paper on AGI Preschool. It > attempts to > define what kind of underlying virtual world infrastructure > an > effective AGI preschool would minimally require. > > thx > Ben G > > > > -- > Ben Goertzel, PhD > CEO, Novamente LLC and Biomind LLC > Director of Research, SIAI > b...@goertzel.org > > "I intend to live forever, or die trying." > -- Groucho Marx ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
