For those who haven't noticed, the range and skill of computer gaming is growing by the day. Most popular deterministic low-move games like checkers or connect four have already been solved, which directly means that computers can play perfectly.
As you step outside those bounds into games with a large number of possible moves, like chess and go, computers have to rely on more or less the same techniques as humans--pattern recognition, looking several moves ahead, and certain predetermined sequences (like opening moves, for example). Add some luck into the mix with games like poker, and perfect play is even less feasible, even if there are less moves to make.
How good can computers get? For move-based, deterministic games, they can theoretically have perfect play, given sufficient time and resources to solve them. With luck involved, they can be perfect within the constraints of the situation, at least, using heuristics to combat unpredictability in the same way we do.
What about games that essentially have heuristics built into them--Jeopardy, crosswords, and puzzles? These games may require not only a large memory but also several guesses and connections, for both human and computer. What is the best way to create fast, reliable algorithms to solve such puzzle games? Matt Ginsberg, the creator of Dr. Fill, argues that in most circumstances, it's best to work inside the realms that computers do best at (and when attempting to beat humans, those realms where they do better). For crosswords, and many other related problems, that realm is memory and search.
But will this be the dominant paradigm for AI and computer decision making? As we learn more and more about neuroscience to find out how and why we are so good at pattern matching and heuristics, might not we be able to grant those same neural algorithms to computers? After all, once a game is solved, it can be programmed perfectly...once our brains are solved, why can't they?
As you step outside those bounds into games with a large number of possible moves, like chess and go, computers have to rely on more or less the same techniques as humans--pattern recognition, looking several moves ahead, and certain predetermined sequences (like opening moves, for example). Add some luck into the mix with games like poker, and perfect play is even less feasible, even if there are less moves to make.
How good can computers get? For move-based, deterministic games, they can theoretically have perfect play, given sufficient time and resources to solve them. With luck involved, they can be perfect within the constraints of the situation, at least, using heuristics to combat unpredictability in the same way we do.
What about games that essentially have heuristics built into them--Jeopardy, crosswords, and puzzles? These games may require not only a large memory but also several guesses and connections, for both human and computer. What is the best way to create fast, reliable algorithms to solve such puzzle games? Matt Ginsberg, the creator of Dr. Fill, argues that in most circumstances, it's best to work inside the realms that computers do best at (and when attempting to beat humans, those realms where they do better). For crosswords, and many other related problems, that realm is memory and search.
"I have a database of pretty much every crossword clue that's been used since around 1990. It's got the body of wikipedia so it looks in there just to see if something comes up, but not a whole lot, it's got a very careful list of words that are probably good crossword words," Ginsberg said. "It just tries to say well I've seen something like this before."Humans are built evolutionarily to be pattern finders, so we have a default advantage in that area. A species without much in the way of fearsome weaponry or agility had better use its brain to recognize the meaning of a rustling bush or a particularly colored berry. However, our memory is fairly limited, mostly in short term but in long term as well; it makes sense, therefore, for computers to focus on memory and search in relevant areas.
But will this be the dominant paradigm for AI and computer decision making? As we learn more and more about neuroscience to find out how and why we are so good at pattern matching and heuristics, might not we be able to grant those same neural algorithms to computers? After all, once a game is solved, it can be programmed perfectly...once our brains are solved, why can't they?
No comments:
Post a Comment