===

In other words, despite the impressive results from many statistical AI systems, a statistical model is nothing magical. Like Google’s capturing of links from web pages, statistical models such as n-grams must be focused on well defined features in consistent pools of data. Then, even once millions of patterns are captured, in any activity with significant variety (like human language) the model must be smoothed to account for the fact that previously-unseen patterns will continually appear. Even once this is accomplished, the model will only be appropriate for limited tasks. Given this, we must bear both the strengths and limitations of statistical approaches in mind. Whether we are listening to intelligence agency officials or singularity-seeking science fiction writers, we should not be fooled by impressive results in certain areas into thinking anyone has developed (or knows how to develop) robust artificial intelligence able to predict, identify, generate, and otherwise stand in for human knowledge of a complex domain.

===

In isolation, it seems that it’s a bit op-ed, with some strong phrases like “utterly inappropriate”, that’s placed between two paragraphs backed up with quotes and calculated reasoning.

In the blog format, it sticks out. In the context of the entire book, it might well be in the right place!

The question is, given it’s not working for you two, why is it not working? One possibility is the one that Ian raises: the blog post format makes the long threads of the book’s argument suffer, even if it’s a good format for looking at the local, specific arguments. Another possibility is that it simply doesn’t belong here (or should be demoted to a note). Hmmm…

As I read it, this section focuses more on some of the inherent limitations; for example, binning together things that aren’t necessarily similar and smoothing out inconsistencies in order to make complex data fit relatively simple statistical models, like modeling sentences as bags of words or n-grams. This is obviously going to cause problems at some point, and might be an argument in favor of more complex models, which is mostly what I read (possibly, misread) this discussion as.

IMO, though, the fundamental issue with what the TIA folks are doing is somewhat different. Although this sort of simplification/smoothing exists, statistical techniques can also give you error estimates: e.g. you can figure out that your naive spam filter, despite not being able to really parse English, nonetheless flags spam with only 0.3% false positives and 3% false negatives, and then decide whether this is good enough for you or not. The main problem with TIA seems to be that they either ignore these error estimates, or decide something is good enough when you or I might disagree that this level of error is really reasonable for a government to use in flagging terrorists.

Incidentally, that sort of critique could probably be made of use of statistics more generally, not only by the government and not only within statistical AI. For example, 95% confidence intervals are taken as “standard” in demonstrating significant effects when doing medical research—despite the fact that this means 1 out of 20 medical studies is likely to contain intervals that don’t correspond to reality, a pretty shockingly high level when dealing with life/death matters.

I think it’s just the final line that *just about* casts what you said in the wrong light.

Google is a good example. Google’s problem is well-defined: find keywords in pages, work out how important they are, match it to user keywords (this is a horrific simplification, but it’s brief ;) ) but if Google wants to be more like a librarian (“find what I mean, not what I say”) it won’t work, because it relies too heavily on keywords.

For Google’s current focus, statistical AI is great, and we see it working very well. For Google’s overall *mission*, statistical AI won’t stretch that far. Continuing your use of Google to show the limitations of statistical AI, but illustrate what it can achieve, would probably get the passage there.

But a lot of people here know *a lot* more than me, so caveat emptor!

I think you and I are in complete agreement about statistical AI. The people we (or at least I) disagree with are the “Total Information Awareness” sort, who are trying to sell politicians and the public on a fantasy of what statistical techniques can manage.

So maybe I should re-cast this paragraph. It’s not actually meant as any sort of attack on statistical AI (which I’m happy to have as part of my daily life). Can you think of a way to make my target clearer?

This chapter is a prime example of what I hope the book can do, which is also what has made writing it a challenge.

Basically, I’m pushing what I can manage — as a thinker and writer — by threading together the past and present, the theoretical and practical, the big issues and the system specifics. Knowing there are readers who like the results is a good feeling.

What people claim for statistical AI varies by person, but the most common view I’ve encountered is actually fairly limited and frank about its limitations—statistical AI says very precise things about very narrow, well-defined problems, so is sort of a bottom-up way of rigorously dealing with problems, starting with the simplest ones. The empirical claim is that it works better than anything else has so far, at least on a variety of common problems, and the theoretical claim is that we have a better understanding of how/why it works or doesn’t work, due to the mathematical underpinnings.

I meant to type this:

I find The Restaurant Game fascinating because it is at a different level of ambition from other work in game AI.

In terms of modeling the background social practices which make everyday stories intelligible (e.g. Schank’s restaurant), there are various levels of simulation: (1) hand-code a few social practices; the system cheats and tells the agents what state the practice is in. (Consequently there is no room for divergence between the different agents’ conception of the state of the practice). (2) data-drive a number of social practices and let the agents themselves work out what state the social practices are in (This is what Rob Zubek does in his dissertation, using markov models to make a best guess about the current state of the practice). (3) Let the system auto-generate practices from statistical analysis of lots of player data. The Restaurant Game is the only project I am aware of which allows for the possibilities of new practices to emerge, which were not defined by the author at design-time.

In terms of modeling the background social practices which make everyday stories intelligible (e.g. Schank’s restaurant), there are various levels of simulation: (1) hand-code a few social practices; the system cheats and tells the agents what state the practice is in. (Consequently there is no room for divergence between the different agents’ conception of the state of the practice). (2) data-drive a number of social practices and let the agents themselves work out what state the social practices are in (This is what Rob Zubek does in his dissertation, using markov models to make a best guess about the current state of the practice). (3) Let the system auto-generate practices from statistical analysis of lots of player data. The Restaurant Game is the only project I am aware of which allows for the possibilities of new practices to emerge, which were not defined by the author at design-time.

Most games ignore social practices altogether. A few games hard-code a couple of social practices.

One person (Rob Zubek) has built a system which
* ignoring social practices entirely (most games)
* hard-coding a couple of social practices
* hand-coding a large number of social practices

…Or, perhaps it is possible to combine statistical and human-authored approaches, to get the best of both worlds. I agree with everything you’ve written, based on what we have published so far, but our long term goal is a semi-automated system where the statistical model forms the scaffolding, facilitating human authoring of the less typical behavior. The Restaurant Game is an ongoing project that will continue to keep me busy for the foreseeable future.

You might be interested in reading pp.8-11:
http://www.media.mit.edu/cogmac/publications/Orkin_JoGD07_inpress.pdf

You note that the player can pay bills with the register or pick it up and steal it. I also wanted to point out that the player has the same interaction options for every object in the restaurant — they can pickup, sit on, bite, and “use” anything. While this opens the door to ridiculous behavior (e.g. biting the waitress, or eating the trash while sitting on the microwave), it also allows players to engage in dramatic improv in ways that we never anticipated. For example, we’ve seen players carry the decorative fruit bowl to the bar, and try to use the blender in response to a customer ordering a margarita.

For example, Bayesian Statistics has been used in AI for decades, so is that traditional or statistical?

Like I said, you’ve probably already defined this earlier, and I’ve just missed it!

I don’t know if they are folding those results back into the search engine yet; but they’re certainly collecting a heck of a lot of useful data.