|
Morgado's Favorite Forum Fascist
Join Date: Oct 2000
Location: Greensboro, NC
|
Interesting Article Regarding Baseball Mogul '07 Tradezilla
I haven't played BM07 much yet, but I just read this article. This was a nice look "under the hood" to me.
Quote:
Tradezilla
April 4, 2006 What is Tradezilla?
by Clay Dreslough, Baseball Mogul Lead Designer
Tradezilla is the name we came up with for the new General Manager artificial intelligence (AI) in Baseball Mogul 2007. While previous versions were pretty good at evaluating trade offers, they still fell short of the intelligence exhibited by the human team owners in Baseball Mogul Online.
As you will see in this article, we tackled this problem using "artificial life": a combination of Darwin's "survival of the fittest" with brute-force computing power. When I thought of "brute-force" and "life", I thought of Godzilla -- so I coined the term Tradezilla for the project.
Player Salary
The core of the General Manager AI is Baseball Mogul's Player Evaluator. Drawing on the research done by Bill James and other sabermetricians, we built algorithms that converted baseball statistics into "wins" for his team.
As an example, according to Bill James' Win Shares methodology, we know that Barry Bonds was worth 18 wins to his team in 2001 (this is the highest value for that year).
The next step to assessing a player's value is to compare him to a "replacement player". A replacement player is an idealized construct that represents any of the various players on the cusp between the minor leagues and a major league roster. These players earn anywhere from about $15,000 per year (a salary in the low minors) to $327,000 (the major league minimum salary in 2006). Although $327,000 is a lot of money, it isn't much compared to the $25,200,000 that Alex Rodriguez is earning.
If Barry Bonds gets injured, you will replace him with the best player you have in the minors. It's the difference between Barry Bonds' performance and that of his replacemement that decides his value.
While the "best minor league player" might vary between teams, it is still a useful concept. If you filled a 25-man roster with "replacement players", you might win about 50 games. So each player would be worth about 2 wins. Barry Bonds was worth 18 wins in 2001 -- so we can also say that he was worth 16 wins "above replacement".
These are the wins that have value -- the wins that we are willing to pay big money to get from a player like Bonds. If you have a team filled with borderline major leaguers, you might win 50 games. To make the playoffs, you probably need to win about 95 games. It's these extra 45 wins that general managers are trying to find every year.
For the purpose of salary, we found that it doesn't matter much where the wins come from. You can pay Alex Rodriguez $25,200,000 to give you 12 extra wins. Or you can pay six separate players $4,200,000 to each give you 2 extra wins. Either way, you're getting 12 more wins than you would without that level of talent.
Player Trade Value
Baseball Mogul's Player Evaluator has undergone many changes over the years. But it is still essentially the process described above. Using the best ideas we can steal from the field of sabermetrics, we assess a player's value in team wins and adjust for the total payroll available to the league.
However, building a General Manager within this framework requires some clever thinking. Evaluating a player's "trade value" is very different from calculating what salary he is worth. This is because the rights to a player also include the burden of that player's salary over the duration of his contract.
Calculating Net Trade Value (First Attempt)
To decide how much a player his worth in a trade, we start with the Player Evaluator and then subtract how much we are paying him (since we could spend this money on other players). If we use the exact same algorithm to evaluate players that we used to determine how much to pay them, we will find that any player that reaches free agency has exactly zero trade value.
For example, if David Ortiz goes on the free agent market and we calculate that he is worth $15,000,000 per year, then one of the computer-controlled GMs will sign him at that salary. But what is Big Papi's trade value if he is already earning $15,000,000 per year? There's no easy answer. If we use the same algorithm we used to decide his salary, then his trade value is exactly zero. This is because his gross value to the team is $15,000,000 per season. Minus the $15,000,000 that we are paying him. And this leaves us with the notion that Ortiz has zero value to the club.
But in real life, this idea doesn't make sense. There are some players like Alex Rodriguez and Mike Lowell that reached a point in their contract where the team felt the player was worth less than what they were being paid. When Texas Rangers owner Tom Hicks woke up and realized that he was paying $25 million per year for A-Rod, and his team was still mired in last place, he knew he was spending too much on one player. Thus he was willing to unload A-Rod to the Yankees and absorb some of his remaining salary. At that point in time, A-Rod, despite being perhaps the most valuable player in the league, actually had negative value as a tradeable commodity, if you factor in his huge contract.
Tradezilla
(New in Baseball Mogul 2007) Tradezilla cont. (page 2 of 3)
Calculating Trade Value (Second Attempt)
So, our first attempt to define trade value was:
(What the player should be paid)
minus
(What the player is getting paid)
| If you're getting $10,000,000 of value from a $5,000,000 player, then you are very happy to have him on the team. And any other team would also be happy to have him. So he has a lot of trade value.
But the above algorithm doesn't always make sense. For starters, most free agents are getting paid exactly what Baseball Mogul thinks they should be getting paid. This may change some if the player's skills improve or decline over the course of a contract. But in general, players that make it to free agency have very little trade value by this method.
This doesn't make sense to baseball fans. We know that players like Hideki Matsui and Mariano Rivera have trade value. Even though they are free agents earning large salaries, you still couldn't pry them away from George Steinbrenner without offering something in return.
The primary reason for this is scarcity. Mariano Rivera might be getting paid $10.5 million. But if you lost Mo and had that money to spend, you couldn't replace him. Even at 36, he has the best cut fastball in the history of the game, and he's still perhaps the league's best closer.
So, the original Baseball Mogul Trade AI incorporated a player's talent and scarcity to assess trade value. This made for a very challenging "artificial GM". But even in Baseball Mogul 2006 there were still some holes that could be exploited. For example, the Trade AI placed value on minor league players that sometimes outweighed their eventual value. Since players develop and different rates and sometimes even change positions, it was hard to find the right balance.
Brute Force Calculations
So, in Baseball Mogul 2006, the Trade AI was essentially the 'player evaluator' with some extra rules added on top. As more people played Baseball Mogul over the last 9 years, they would send us their feedback and we would incorporate it into the game.
Some customer might tell us that Baseball Mogul puts too much value on strong defensive middle infielders. In other words, after playing the game for hundreds of simulated seasons, they would report that it was possible to trade a marginal slick-fielding weak-hitting shortstop for a much more valuable players. Others would report
We took all of this feedback and added it to the game. Essentially, we were using brute force testing to improve our algorithm. Thousands of customers would play the game and report things that didn't seem realistic. And we would add a rule to the algorithm.
It's not an easy process. We're trying to make a computer as smart as a human General Manager. And it's just one part of a much more complex baseball computer game. User feedback was improving the AI. But only a little bit at a time.
Tradezilla cont. (page 3 of 3)
A Different Kind of Brute Force
Instead of constantly updating our AI in response to user feedback, I tried to think about a better way to advance the artificial intelligence.
The main goal of the system is to build an intelligent "Artificial General Manager". One that can win baseball games and championships within the world of Baseball Mogul -- an extremely detailed simulation of Major League Baseball.
We know how "real" General Managers are made. Millions of years of evolution have created the human brain. Like the user feedback we've already been using, evolution is a "brute force" method. It creates billions of organisms, each with different traits, and only the best ones survive.
Evolving a General Manager
I decided to use this system for creating an intelligent General Manager (or, as it turns out, a group of intelligent General Managers, each fine tuned to their own circumstances).
Because Baseball Mogul can simulate an entire season in less than 30 seconds, I realized that by leaving my computer running for a few days I could simulate over 10,000 seasons -- the equivalent of thousands of entire careers for a real-life a General Manager.
In the simulation, one General Manager career correlates to one generation of natural selection. We started the simulation with 30 randomly generated General Managers. Each General Manager was defined in 28 dimensions, from the value they placed on young relief pitchers to whether they prefered drafting high school or college players.
All of these 28 dimensions interacted with each other. So one General Manager might have a penchant for drafting left-handed college pitchers with great "control" ratings while choosing to trade for power-hitting veteran outfielders. We ended up with over 2.6 x 1064 different possible General Manager profiles.
In addition, these profiles also incorporated how General Managers worked within the constraints of their organization. One profile might define a GM that always paid top dollar for the newest free agent, while another profile describes a GM that only does this when he has a surplus in his budget.
We then began simulating our artificial baseball world. Every 5-10 seasons we would stop and measure every GM's success, based on games and championships won. The bottom 80% of GMs would be fired and the top 20% would retire. However, the top 20% also became the "parents" of the next generation of GMs. For each surviving GM, we randomly altered his profile, like the mutations that power evolution.
This process continued for 11,000 simulated baseball seasons -- enough to take us past the year 13000. We then filtered through the successful profiles, compared them to the other successful profiles, and essentially distilled a set of unique profiles that were the most successful.
We copied these profiles back into Baseball Mogul 2007, and they are now used as the actual Artificial General Managers that you must haggle with on a daily basis in order to make trades and improve your team.
|
One obvious question: does this mean, then, that the GM's are each notably unique in the game, or does it mean that, due to the mutations and evolution that the surviving ones are all very similar?
__________________
The media don't understand the kinds of problems and pressures 54 million come wit'!
|
05-01-2006, 06:44 AM
