Category: Artificial Intelligence
I’m not a novelist. I can write, and with sufficient editing, I can even write passable English text, but it’s not really something that I’ve put a lot of effort into, and so not something that I’m good at. My real time investment in writing has been writing software, which is read only by compilers and interpreters, and so what it lacks in excitement, narrative, and plausible dialog, it makes up for in conciseness and precision.
My ideas for National Novel Writing Month (November, for those as don’t know) then, are coding a program that writes novels, and coding a program that takes novels as inputs and generates interactive fiction (text adventure games) from them. Both of these are problems with infinite hair, and are arguably AI-Hard problems (that is, problems whose solution is on par, difficulty-wise, with creating a human-level general-purpose AI). On the other hand, writing a good novel is probably also quite hard.
I think the most reasonable approach for the novel generator would be a recursively-defined novel description language, which selects from tropes and plot stubs, generates characters, and so forth, based on relatively simple rules. The complexity would come from applying the rules over and over, so a simple quest to throw a ring into a volcano grows branches on branches on branches until it is One Damn Thing After Another Until All The Orcs Are Dead. The goal of the program would be to use generative content and emergent behavior to do most of the writing, and leave me to fill out the turns of phrase and details (or generate them a la Dwarf Fortress, which menaces with spikes of ivory). Done badly, this would read like a Mad Lib. Done well, it would read like a Mad Lib filled out by people who don’t say “dongs” every time they are asked for a noun.
Making interactive fiction (IF) out of novels would be substantially harder. The novel parser would have to read English, which is actually quite a trick. English has multiple words for one meaning and multiple meanings for one word, highly flexible structure, and counts on the reader to sort it all out. On top of that, most of the awesome tricks one can pull in English are more a matter of exploiting shared cultural context with your reader than they are particular sequences of words. If I wrote such a parser in one month, or at all, a lot of linguistics researchers would be out of work.
Assume I went for one tiny part of the problem: identifying the locations in the novel. The same place might be described as “where that party was”, “Joe’s house”, “the darkened house”, and “a pit of iniquity”. Only the events of the novel link them, and so the program would have to determine that these totally different words referred to the same place. The best approximation I could likely come up with is identifying all the things in the novel that sound like places, and then performing some sort of clustering based on what words are mentioned close to mentions of those places. This would likely lead to a bunch of spurious places getting generated, and real places getting overlooked. There is an entire company, called Metacarta, that did this sort of analysis on much more constrained data sets, and even then it was a difficult problem for a team of people who were likely smarter than me.
However, doing a good job of adapting novels to interactive fiction might not be the best approach. It might be better to get a rough cut of the software together to do anything at all, and gradually improve it until it writes things that are playable curios, rather than detailed simulations of well-loved novels. It wouldn’t be a matter of playing through “A Game of Thrones” so much as it would be “poking around in a demented dreamscape based loosely on ‘A Game of Thrones'”.
This is actually related to another idea that I had, which is sort of a rails shooter based on the consequences of shooting things. You play through the game, riding the rails and shooting enemies of varying levels of craftiness and menace. When you reach the end of the level, you just loop through it again, passing the bodies of everything you killed, and getting another shot at everything you missed. This repeats until you have killed everything in the game world, whereupon you continue to loop, passing through scenes of slaughter as the heroic music fades and is replaced with silence and the buzzing of flies. Perhaps, if you let it run long enough, the dead bodies would rot to skeletons. Now, of course, I’ve spoiled it for you, but I’ll probably never get around to writing it, so at least you’ve had the idea.
I finished writing the chatbot that I was working on. It consists of a set of scripts to prepare the data, and another script that listens for incoming messages and responds. You can get the code and an overview of how it works here.
Obviously, I’m not publishing my chat logs. Use your own. It is designed to work with Pidgin’s HTML-like format for chat logs, but it could be modified to work on almost any corpus. I really should clean up things like the string cleaning routines, but it worked for class, and that’s what actually matters.
I have been working, on and off, on a computer vision application that will recognize a card from a certain game in an image, figure out what the card is, and add it to a database. I had some luck in earlier versions of the code looking for template matches to try to find distinctive card elements, but that fails if the card is scaled or skewed, and it rapidly becomes too processor-heavy if there are many templates to match. Recently, at work, I have had even more opportunity to play with OpenCV (a computer vision library), and have found a few blogs and tricks that might help me out.
The first blog shows how to pick out a Sudoku puzzle from a picture. The most important part is finding the corners of the puzzle, as after that, it can be mapped to a square with a perspective transform. I can do a similar trick, only I’ll be mapping to a rectangle. Since corner-finding is kind of scale-invariant (the corner of something is a corner at any scale), this will let me track a card pretty easily.
I think that I can actually use OpenCV’s contour finding to get most of the edges of the card, and then the Hough transform to get the corner points. I may even be able to get away with using just contour finding, getting the bounding rectangle of each contour, and checking that it has something like the proper aspect ratio. This will work in the presence of cards that are rotated, but fails on perspective-related skewing.
This StackOverflow post has a nice approach to getting the corners of a rectangle that has some rotation and perspective skew.
Once I have the card located, I’m going to throw a cascade of classifiers at it and try something like AdaBoost to get a good idea of which card it is. Some of the classifiers are simple, things like determining the color of the front of the card. Others may actually pull in a bit of OCR or template-based image recognition on (tiny) subsections of the card. Since I will actually know the card border at this point, I can scale the templates to match the card, and get solid matches fast.
Computer vision is hard.
I have code written that detects edges in a video image, picks pairs of edges that are both within a certain ratio of lengths relative to each other and within a specific angle of each other. I’ll post the code up once I have pushing to my github account set up correctly. It’s actually pretty poor at locating cards in a video image, but it is a demo of how to use some of the UI and feature recognition stuff from OpenCV.
From here, I have a few ideas about how to proceed. The first is to have the program display a rectangle and ask the user to align the card to that rectangle. This means the program will not have to find the card in the image, and can focus on recognizing parts of the image. This will use template-based recognition to pick out mana symbols, and should be pretty simple. The classifier that tries to pick out colors will be even easier, as it will just select a sample region, blur it, and match it to precomputed colors. This can be calibrated for specific lighting conditions for each run.
An even less hands-on approach was suggested by my co-worker Eric. The cards could be displayed by a machine that essentially deals one card at a time. The machine would have a magazine of cards, and would photograph a card, deal it into a hopper, photograph the next card, and so forth, until it was empty. This would have the problem that it wouldn’t be something anyone could download and use, as it would require a special card-handling machine, but the software end would be compatible with the user-based registration approach described above, so someone with a lot of patience could use the manual method.
Another approach would be to throw all the cards on a flatbed scanner (possibly with a funny-colored (e.g. red) bed lid, do template matching to locate card borders, and then segment the images based on that. An ADF scanner with a business card attachment could probably also make short work of a modestly-sized set of cards.
I used to play the collectible card game (CCG) Magic:The Gathering. Like most CCGs, Magic has a large set of different cards that players can use to build a set for playing games. This is both fun, as it means new cards will allow new play types and strategies, and annoying because of the artificial rarity of some of the cards. I don’t have a lot of people to play with, so I am planning to sell my cards.
I will probably make more money selling the cards as individual cards (“singles”) than I would get by selling the whole set. However, that means that I need to know how many of each card I have. Given that I probably have upwards of 8,000 cards, I don’t want to sit down and type in the name of each card. It would be better if I could have a computer program do it for me, so I’m working on writing one. The rest of this article uses jargon from Magic and computer vision, so it may be a little incomprehensible to people who are not Magic-playing computer vision nerds.
The program will take an image using a web cam and look for two straight edges, probably using some form of edge detection or a Hough transform. Once it has the edges, it will look for two edges whose ratio of lengths is the same as a Magic card. The edges must share an endpoint, so that the system can tell they are the edges of the same object. The area inside the rectangle that has those lines as its edges is the card.
Once the card is detected, the simplest thing to do is to match the card image against card images stored in a massive database of all card images. Unfortunately, there are over 11,000 unique cards (11,458 as of Feb 2009), which would make for a processor-intensive comparison process.
My plan to circumvent this is to have the program get the casting cost of the card by using processing techniques similar to face detection. The most useful technology to detect mana symbols is probably feature-based template matching. Feature-based template matching allows the computer to pick out a region of a picture that matches another, smaller image, even in the presence of some distortion. Mana symbols haven’t changed significantly since the development of the game, so they should be easy to pick out.
I can also get the color of the card by selecting a region of border, blurring it to remove any texture, and comparing the color to a set of swatches. I’ve done this sort of comparison before, by measuring distance in RGB color space, and it can be done quickly and effectively. The main possible pitfall is poor lighting shifting the color of the object, but I can at least arrive at a probabilistic result based on the RGB distance. Combining the estimated color of the card and the casting cost will allow me to significantly reduce the set of possible pictures that the card image needs to be matched against.
There is also the question of building the database of card images, but I believe I can do that by screen-scraping the web site of the company that makes Magic Cards. I won’t be able to distribute the database with my program, as it will contain copies of copyrighted data, but I can distribute the screen-scraping script.
I may also be able to recognize features like card text, but that will rely on very good lighting and very good cameras. I would prefer that this program work with a moderately high-quality webcam, so that it will be useful to people other than me.
The recognizer will try to build a list of cards that it thinks matches, ordered by the confidence of the match. If the user rejects the first card, it will present the next, until it runs out of guesses. If the use accepts the guess, the recognizer will add that card to a database of all the cards the user owns. In this manner, the user can build a database of cards by simply showing the cards to a computer.