Word Freak : Heartbreak, Triumph, Genius, and Obsession in the World of Competitive Scrabble Players (9780547524313)

by Fatsis, Stefan

  19. 1501

  ON THE THURSDAY after my triumph at Bird-in-Hand, I arrive at the club for the regular weekly session and hand G.I. Joel the $11 fee.

  “How are things?” I ask.

  “The same,” Joel says, checking off my name on the club master list and writing it and my old rating on an orange slip on which I will record the evening’s results.

  “Good,” I reply. “I guess.”

  “Not if you’re me.”

  I smile and take a couple of steps into the room.

  “Oh, Stefan?” Joel says.

  “Yeah?”

  “Way to go.”

  At tournaments, I always feel as if I’m playing for Joel, as if he’s a coach who lives for his team, or a kid in a hospital bed for whom I’ve promised to hit a home run. Tournament results are posted on CGP, so club members know that I was a winner, and inside the playing room I’m greeted with everything but ticker tape. “Kickin’ some tush!” Jeremy Frank says, and slaps me on the back. “Way to go, kid!” Sal Piro gushes. Even the petulant Mark Berg offers congratulations.

  In the second game of the evening, against a weak opponent, I open with a deliberate phony, MEAOW*. On the next turn, she takes the bait, pluralizing the fake word, and I challenge that off of the board and gain a turn. It’s a devious tactic that requires you to have more knowledge than your opponent. It’s also praiseworthy. At the next table, one of the old-timers watches the sequence. “You’ve become one of us,” she says.

  During a game, there is precious little time to perform the complex calculus involved in making a play: assessing the seven letters on your rack, analyzing the state of the board, determining the possible moves, and deciding which among them maximizes your chances of winning and minimizes opportunities for your opponent. Instinct rules, but in this case instinct is acquired, the product of hundreds of hours of play and study. Even as transparent a ploy as playing a deliberate phony to lure your opponent into tacking an S onto it is a product of that learning.

  Scrabble theory grew up in the early 1990s. The computer program Maven was accepted as a force in deconstructing the game. The newsletters Medleys and Rack Your Brain laid new mathematical and strategic roads. They were deep geek. Rack Your Brain devoted an entire issue, ten pages, to analyzing a 5-point endgame play. There was even a Scrabble humor ’zine, Moxbib, named after a word coined by Joel Wapnick to denote a ridiculously absurd phony. (Moxbib included features such as a list of the two-letter combinations not found in the OSPD and a TV guide with shows like Scrabble, She Wrote and 700 Club, on which two players “relive their 700-point games of a decade ago.”)

  Maven was the key to theoretical study. It asked which was more powerful: the brain, with its intuitive muscle but chaotic operation, or the computer program, with its orderly function but no intuition. Of course it wasn’t a new question. Games have been considered an ideal model for studying artificial intelligence since the late 1940s, when the first paper on programming a computer to play chess was written by a researcher at Bell Telephone Laboratories in New Jersey. Early chess move analyzers—with names like Turochamp and Machiavelli—simulated the play of a computer one move ahead, or one ply, by attaching scores to each possible position and making a move based on the best possible score. Later, a program called SOMA attached material values to each piece—the queen 9 points, rooks 5, knights and bishops 3, pawns 1. SOMA measured mobility based on which squares were being attacked, and created values for exchange sequences. Initially, computers didn’t do the actual work; the scientists performed the calculations, known as simulations, by hand.

  In 1983, during a summer internship at IBM’s research center in Yorktown Heights, New York, later the home of the famous Deep Blue chess program, a Harvard undergraduate named Brian Sheppard came across an article about a Scrabble computer program in a journal on artificial intelligence. The article said the program averaged 19 points per turn. Sheppard considered himself a terrible Scrabble player, but he felt he could do better than that; he had heard that experts averaged well over 20 points. That summer, he wrote a simple program and typed in about twenty-five thousand words from the OSPD—every word of five or fewer letters plus the J, Q, X, and Z words. The program averaged about 23 points per move. Brian went back to school.

  Three years later, bored with his job, Sheppard picked up the Scrabble program again. This time, he manually entered every OSPD word into the program. With little built-in strategy, the revised program averaged 30 points per turn. Sheppard named the program Maven after a computer he once had worked on, only later learning that the word itself meant “expert.”

  To make Maven play smarter, Sheppard needed an “evaluator” so the computer could make informed decisions, rather than just playing words for the highest possible score, which had resulted in the program sticking itself with terrible leaves. Sheppard had the computer play thousands of games against itself and he recorded the results, giving the letters point values, as the chess programs did. For instance, Sheppard kept track of how many extra points the program scored when it kept an S, and that became the value of the S: + 8 points. An E was worth +4 points, a blank +25, the Q—13. Two I’s on a rack was—10, while UU was—12. Sheppard rated one combination, QU, which turned out to be neutral. He assigned a negative value to opening a triple-word-score line, to prevent the computer from doing that willy-nilly. And he assigned values based on the balance between vowels and consonants left on the rack.

  “I figured that on each turn three things change,” Sheppard tells me when I call him. “You score points, you change the tiles on your rack, and you change position.” So, in making a play, Maven determined that the value of a move equaled the score of the play plus the value of the rack leave plus the board position. The board was neutral in almost all circumstances because it was a resource shared by both players. Maven computed a value for each possible play in a turn and picked the best one.

  Then Sheppard got in touch with two Scrabble experts in Boston, who coincidentally had been testing another computer program. Sheppard invited them to play Maven. “It absolutely kicked their butts,” he says. Maven wasn’t perfect—it screwed up endgames, it had dictionary glitches, it didn’t assess positions optimally, and it was often needlessly offensive. But Sheppard reasoned it was ready to take on the pros. Another Boston expert was testing his program in a tournament and invited Maven to enter, too. Against the likes of Joe Edley and Bob Felt, Maven finished with an 8–2 record, good for second place.

  Though he had never played seriously—he knew so little about Scrabble that in writing his code he called racks “trays”—Sheppard had developed a program that could beat the very best. And in the process he disproved much of the conventional wisdom about the game. By letting the computer play itself, Sheppard determined that it wasn’t worth sacrificing points to avoid placing a vowel next to a double-letter-score square to prevent a big comeback play. The notion of tile turnover—moving as many tiles as possible to shoot for the blanks and S’s—also proved mathematically insupportable; turning over more tiles encouraged the computer to keep more bad tiles, hurting its longer-run chances. General theory held that opening the board was bad because it gave the opponent first crack at new bingo lines; Maven’s play determined there was no penalty for openness, except on the triple-word-score lines.

  Sheppard perfected the program’s dictionary so Maven wouldn’t play phonies or challenge good words. He improved its ability to play endgames so it could block plays and not get stuck with tiles. Then he met Ron Tiekert, who told him about a test he had done involving a single rack of tiles.

  At the time, the mid-1980s, only a few top players were combining strategy and mathematics. Tiekert, for instance, had tried to figure out whether A or I was the “better” letter. Starting with an A he drew six additional tiles and made a play. He returned all seven tiles to the bag, removed an I, and drew six tiles to go with it. Tiekert repeated the test over and over to see which yielded a higher score or
fruitful leave most often. He did the same thing with a few two-tile combinations, and then applied the same logic to full racks. He wanted to figure out the best play to open a game.

  Opening-rack analysis was pure, uncomplicated by board position or score. One of Tiekert’s racks intrigued him: AAADERW. There were three obvious plays—AWARD, AWARE, and WARED, scoring from 18 to 26 points. But Ron had a hunch that AWA (Scottish for away) was the right move. It went completely against conventional wisdom. AWA scored just 12 points, and it moved two fewer tiles than the other choices, so by the popular thinking of the time it had to be a loser. But Ron realized that its leave of ADER was likely to produce a bingo on the next turn or two. Also, AWA didn’t allow one’s opponent access to double-word-score squares, and it couldn’t be made plural with an S as AWARD could.

  Ron had heard about backgammon experts who would spend hours replaying the same position to determine the “right” move. Ron figured such a test would work in Scrabble. At a table in his room in the Commander, the SRO where he lived, Ron pulled AAADERW from the bag. First he placed AWA on the board. Then he drew a rack of tiles for a mythical opponent and made the best play, recording the score. Then he replenished his own rack with three letters and made the best play, recording that score. Next, he placed AWARD on the board. He took the same rack he already had pulled for his opponent and made the best play. Then he drew two more tiles to go along with the three he had drawn to refill his rack after AWA and made the best play. He did the same thing with AWARE. At the end of the three turns, he added up the scores, giving bonus points based on the quality of the leaves.

  Ron repeated the exercise one hundred times and totaled the results. Just as he suspected, AWA resulted in the highest average score after three turns. That didn’t mean that AWA was guaranteed to produce a victory. It simply indicated which move was better from a probabilistic perspective. Ron submitted the rack for discussion in an expert newsletter and waited to see the response. No one else picked AWA, and most didn’t even mention it as a possibility.

  If a computer could be harnessed to perform such a test, Ron surmised, it would help players better understand their decisions. So when Ron met Sheppard, he told him about his manual test and suggested that Maven tackle such problems. Sheppard wrote a program that ran a thousand trials of each rack position, and it verified Ron’s finding that AWA was the superior play. Sheppard added the “simulation” function, or “sim,” to Maven.

  By the early nineties, more and more top experts were relying on Maven. It helped them to see when and where they made mistakes. “Everybody has leakage,” Charlie Carroll, a Minnesota computer programmer and a top player at the time, tells me. “Where are you losing points? Are you missing bingos? Are you missing overlaps? Is it in the endgame? This was the perfect tool.” Joe Edley—who was contemptuous after losing to Maven in the program’s tournament debut—eventually required that every game analyzed in the National Scrabble Association newsletter be checked against the program. (Hasbro eventually hired Sheppard and bought Maven to create a Scrabble CD-ROM, which is considered stronger than any human player; it beat Matt and G.I. Joel six games to three in a match arranged by The New York Times Magazine in 1998 for a story on Scrabble’s fiftieth anniversary.)

  Brian Sheppard and Medleys editor Nick Ballard began pushing mathematical theories. Their main contention was that Maven’s tile values could be considered not only by a computer but a player over the board. Maven ranked the tiles from best to worst, like so: blank, S, E, X, Z, R, A, H, N, C, D, M, T, I, J, K, L, P, O, Y, F, B, G, W, U, V, Q. Scrabble is a zero-sum game, Ballard and Sheppard theorized, with the value of the hundred tiles totaling zero. Players could perform running calculations of the value of plays and of the bag during a game. The long-term goal was to give everything a numerical value so that evaluating candidate moves would become mechanical, requiring players just to add up a bunch of numbers.

  Some experts scoffed, others marveled. But once Maven began simulating plays for them using the tile values, lightbulbs clicked on. Not only was there computer analysis to back up moves, but Maven could help answer questions. What letters were better to keep or trade? At which point in the game was creating board volatility best? What was the cost of particular plays and counterplays? “It was a puzzle,” Carroll says. “It was different than chess, which has been analyzed to death for hundreds of years. You got to figure out stuff that no one had ever figured out before.”

  There are few theoretical breakthroughs anymore. The Scrabble debate has migrated on-line to CGP, which, like any chat room, is polluted by sniping, bad jokes, self-aggrandizing commentary, and inconsequential banter; I automatically delete three-quarters of the posts. But when board positions are posted, the quality improves. Computers are consulted to conduct simulations consisting of thousands of trials, known as iterations. Some players love sims, a few dismiss them as irrelevant; Marlon says either you win a game or you don’t—every game is different—and a sim doesn’t influence that. Too many players, especially the programmers, he says, consider Maven’s judgments to be gospel. “Sim is doo-doo,” Marlon says.

  But sims reveal not whether a play wins or loses a specific game but whether the play yields a higher probability of winning over all other moves. “Humans can’t do sixty-five hundred iterations in a twenty-five-minute game, so functionally you have to go with your intuition,” Bob Felt says. “At end, intuition is distilled experience, and simulation provides reams of experience. The point of doing simulations is not to find out what play you should have played, but to change your thinking so you are more likely to make the play you should make in the future.”

  So while I gape at the tileheads, I also try to think more like them. In a tournament game, I draw an opening rack of CEEGPP?. I play PEP. Afterward, I wonder whether PEP was best and ask G.I. Joel for help. After 6,910 three-ply iterations, Maven says PEG wins 56.1 percent of games, followed by PEC, CEP, and then PEP. I picked the fourth-best move, which wins just 47.5 percent of the time. (Joel tells me that the G doesn’t naturally blend well with either the C or the P. The best consonant combo among these three letters is CP, he says, and the C is a keeper because it’s a good bingo tile.)

  Poring over the e-mails and theory articles and old Scrabble games played by the masters seems to be changing how I think. Chess players learn by committing great games to memory and dissecting the positions of Fischer, Morphy, Capablanca, and Botvinnik. Substitute the names Edley, Cappelletto, Sherman, and Gibson, and Scrabble is no different. Over the board, there is no way I can calculate whether the “equity” of a particular leave is—4.5 points or—3.5 points, or that the T has a value of—0.8 points while the N is—0.2. But I can understand why it matters, and use the information to think more logically, rationally, and mathematically—more deeply.

  It’s part of the organic process of getting better at something. The payoff from studying words is obvious; I recognize more words and play them. The intuitive changes occurring in my brain are subtler: having a better spatial sense of the board, spotting bingos instantly, reaching into a full bag and knowing I have drawn exactly seven tiles, instinctively knowing that one word is better than another at a given point in a game. To paraphrase Supreme Court Justice Potter Stewart’s comment on pornography, I can’t yet define a good Scrabble play, but I’m starting to know one when I see one.

  I don’t sleep well at Scrabble tournaments. The soft hotel mattresses with their down-free pillows and synthetic blankets are uncomfortable. And there’s always a roommate, a snoring Marlon or a jittery Matt. I roomed solo at my first few events, but began feeling foolish, the journalist solidifying his outsider identity, refusing to be one of the crowd, and wasting his money; the extra hundred bucks per tournament seems needlessly extravagant, especially with Matt or Marlon borrowing bus fare or cadging meal money.

  The twenty-game Eastern Scrabble Championships over Presidents’ Day weekend doesn’t promise to be a restful event. A blizzard has dumped e
ight inches on Danbury and paralyzed highways around the east. Players are straggling in late, if at all. Six players get into car accidents en route; one guy totals his vehicle, hitting his head on the windshield, and then simply abandons the car and takes a train to town. (That’s nothing. A tornado once destroyed part of an Ohio hotel that was staging a tournament; play continued in the unaffected section.) A total of 113 players finally make it.

  I’ve talked myself into believing that this event will define my progress. It’s a strong field. But my rating went to 1501 after Bird-in-Hand and now is up to 1524. I’ve been studying hard and have been exercising regularly to build up my energy level; upon arrival at the Inn at Ethan Allen, I head to the tiny gym and ride the stationary bike (thirty-five minutes), then run on the treadmill (fifteen minutes) before the evening’s opening rounds. I stock up at the local Super Stop & Shop: a half dozen bananas, three oranges, two bottles of water, five Clif Bars, three Think! bars, one Fresh Samantha Super Juice with Echinacea. I’m prepared, for Scrabble or a few days in a fallout shelter.

  But play starts more than an hour late because of the snow, and Friday night’s three games don’t end until well after midnight. I go 21 but am exhausted. The clock in our hotel room is an hour fast, which neither Matt nor I realizes until the following morning. So when I climb into bed, I think it’s 2:00 A.M. Panic sets in: I’m overtired, I’m not going to get enough sleep, I’m not going to be alert in the morning, I’m doomed to play poorly. When I awaken from a state of borderline consciousness after a fitful night, I think it’s 9:00, but it’s only 8:00.

  Nonetheless, by the end of Saturday, I have a 7–3 record and am in third place out of the thirty-four players in Division 2. I eschew after-hours Scrabble in the game room, but make the mistake of checking the name of my first Sunday-morning opponent. It’s Amit Chakrabarti, who is leading the field with a 9–1 record in his very first tournament. A native of India, Amit is an experienced SOWPODS and Internet player; at the suggestion of Bob Felt and G.I. Joel, who have played against him on-line, Amit was placed in Division 2 by Ron Tiekert, who runs Danbury. Amit is a prototypical expert: a computer science Ph.D. candidate at Princeton whose resume (I look it up later on his Web site) includes a paper he cowrote titled “A Lower Bound on the Complexity of Approximate Nearest Neighbor Searching on the Hamming Cube.”