Stan Jedrus


"More Intelligent" Artificial Intelligence

Chances are you have seen "2001: A Space Odyssey". If you haven't, then perhaps you've seen something from "Star Trek: Next Generation" or "Star Trek: Voyager" series. All of those productions have one thing in common: they try to picture a little more intelligent "artificial intelligence" - concious, responsible for their actions, creative to some degree (versus the commonly known: full of preprogrammed reactions, based purely on statistical analysis of previous inputs).

Of course, there have been a lot of efforts to "spice up" the current AI to the point where it might pass a Turing Test. I think one of the most relevant developments in the area has been done by these folks. Their ACT-R model and software implementation is probably the most advanced published idea about AI so far. And at the same time, it's hard not to notice how far behind this model still is. The model itself has hardly changed since it's original publication over 30 years ago. And while the implementations start to have more and more options, they don't effectively push us anywhere close to what Stanley Kubrick dreamt about almost 40 years ago.

I'd love to be the person to develop the real "HAL 9000". At the same time, I realize that this is nearly impossible at the moment. Therefore I decided to develop something a little smaller - maybe "HAL 1". I think I'm pretty far on the way. I hope to have the prototype completed by this summer. Either way, if I manage to finish this prototype, it will probably change the way we think about computers. If it's successful, maybe I'll get to make a "9000 version" as well.


Computer Networks

Computer Networks have been my long time interest. However, after reading paper after paper after paper published on the subject, and reviewing many for upcoming conferences and journals, I've come to the conclusion that the subject is pretty much closed. The most recent research seems to go in either of two directions: marginal improvements that have exactly zero impact on practical applications (because nobody is going to bother considering the existing installed base), or abstract ideas that are comparable to "string theory" - impossible to prove or disprove, while very convenient to publish a bunch of papers and maybe even get some grant money. It seems like genuine impact of research on networking ceased about the same time when the globe has been wrapped tightly in the network of superfast optical fiber links. The capacity of those links in most cases greatly exceeds the current needs so the demand for research on any subject related to telecommunication simply doesn't exist.

At the moment, computer networks suffer from lack of applications that might use them. My theory on the current problems of telecommunications industry is the following: people are going to only use so much of network capacity that causes the latency to hover around the threshold of acceptable level. But since a lot of the latency can't be really reduced because of the physical limitations, and "acceptable level" for many of the existing applications means rather low than high latency, a lot of network capacity will just sit there unused. The networks "seem" to be empty, but really they are saturated from the point of view of the user, who often can't tolerate latency fluctuations traditionally associated with higher network loads. While it is technically possible to separate different types of traffic, effectively giving some QoS at least to one type or another, in practice it's not that useful idea since it's rather easy to "disguise" one type of traffic as another (at least in IP networks.)

The real culprit of the situation lies in the very nature of TCP/IP - no guarantees, best effort, packet switching. The architecture itself imposes limitations on what can be done with it. While there has been a lot of work done on the subject of QoS in TCP/IP, the reality is rather stark - those nifty features like MPLS, Diffserv, and so on, are not really that useful to provide actual QoS for users, and even further from actually being used for that purpose by the ISPs. The question here is simple and as old as ATM networks: who is supposed to decide what quality of service particular application requires? The ISPs have always said that the users will decide - and that has been first implemented as "per connection" idea (ATM), and later as "per end-user" (like in MPLS, or Diffserv.) The former has just a simple disadvantage of not being scalable so it was very quickly abandoned, the latter, while scalable, it forces users to actually actively negotiate particular QoS requirements with the ISPs in advance. That situation forces the users to know a lot about what their network needs might be, which is pretty much impossible to realistically say about most users connected to the Internet, at the moment. Also, another problem here is that ISPs don't really deliver the product the users are interested in - ISPs are just the middle-man. So, ultimately, even if say user A pays for a better QoS access, they have no guarantee to actually receive that QoS because the other end of the connection might be less inclined to throw away money on something as volatile as QoS. In practice, QoS is really out of control of particular ISP - they control only a very small portion of the whole network, and even if both ends of the connection are connected to the same ISP, they might be paying for different QoS, so one of them will get the shorter end of the deal.

The ISPs are very well aware of this whole problem, so if you get any type of general connectivity to the Internet, the first thing you're going to hear from the ISP is that they do not guarantee the network speed - they just give you the access connection of particular speed. While this is generally accepted because the bandwidth costs are relatively low at the moment, it wouldn't be so easily acceptable if people were actually contracting ISPs to deliver particular level of QoS. So, while technically all of those nice ideas might exist, in practice, most likely we're not going to see anything better than the good old "best effort". Which means that most of the research in the IP/QoS domain is quite meaningless.


Time Series

My interest in accurate modeling of time series has started over ten years ago when I worked on my PhD thesis. The main fruit of that interest was the development of various kinds of multifractal models for the network traffic. Technically, time series is the same thing no matter whether it's a network traffic, or temperature on the surface of Pluto, so the models are actually rather universal. Again, although I'm still interested in this area (mostly for the purpose of modeling stock and index prices), I haven't done much of creative research in this area in quite a while. All I can say about my models of stock and index prices is that it seems like it won't really make any big difference to invest in stock for the next 20-30 years compared to investing in just about anything else that can give a reasonable interest rate. Unless you are into speculation, that is - which is usually equivalent to playing Black Jack in Las Vegas (if you have been "winning" - remember that some people do win in Las Vegas, as well.)

Contact Me | ©2006 Stanislaw Jedrus