I’ve recently been following the progress of computer AI in playing poker. One area of interest is that the AI responds to bluffing. Bluffing is a major aspect of poker that makes it interesting from an AI and gaming perspective.
Around 1977-1978, I programmed a chess playing opponent in 4k of memory on my first computer, a Processor Technology Sol-20 based on an Intel 8080 cpu.
I think the 8080 ran at 1 MHz to 2 MHz, so about 1 million to 2 million instructions per second. Today’s Intel Core i5 processors run closer to 3 GHz — about 3 billion instructions per second — and that doesn’t even take into account multiple cores for parallel processing. That’s 1.5 thousand to 3 thousand times faster than my 8080-based computer. So you can see that at the speed of the 8080 an AI couldn’t depend entirely on cpu-devouring depth searches and tree pruning algorithms to determine its next move. That’s why I added a bluffing component.
I don’t remember if my computer had 8k or 16k or memory, but just for reference today’s phones with 16GB of memory have 16 million times more memory — since 16GB = 16,000,000k. Okay, so with just 4k of memory allocated to my chess game to handle the display, game logic, input, output, and AI, I was very limited to what I could do with bluffing. Actually, the bluffing component was coded so simply that it was almost a random move injector. But I believe it was that aspect of Fischer — the temporary name I gave to my chess program — that allowed it to sometimes compete with other chess programs at the time. From time to time it would make a bold move — a leap beyond it’s ability to just search for the best next move — effectively bluffing that it had a plan that the other program could not discover in a depth search of the possibilities.
So bluffing can be useful even in non-poker games, although that only works until the game has been “solved” by computers. There are games like Checkers that have been solved by computers, meaning that the entire game is known from the start. Even some games of Poker, e.g. Head’s Up Limit Hold’em, are largely solved.
In poker, bluffing is not a solution for an AI, but rather a necessary tool. It’s built into the game of poker. So far, looking at comments on Reddit, it appears that the best poker AI can play with the best poker players. Good luck bluffing your way through the tournament. Also, this might be the beginning of the end for internet poker.
If you have further interest in this subject, you might also like this recent research paper.