Monday, September 27, 2010

Class notes: 9/27/10

Tournament format:
Warning: If your player errors out during the tournament, it will be removed from the rest of the tournament

Do we include old players?
Neha: Include just G1
KR: Let's have 2 classes of tournaments - one with just new players, one with both
Daniel: Some have bugs
KR: OK, so we will use G1 as a representative of last year's players
KR: Note that we are optimizing for high score, rather than ranking

KR: Any ideas for a new tournament, not one that we did last year?
Nitin: Bring in more of LessStingyPlayer
DFed: Using LessStringyPlayer will hurt conservative players
KR: Is it interesting to run with 7 hidden letters?
Elba: Yeah, it's pretty interesting. People still get 1 letter. There is still some play
Nitin: With 7 hidden letters, LessStingPlayer should be there.
Yang Jian: 7 hidden letters means that we depend a lot more on luck
Eva: With 7 hidden letters, we should have more rounds
Daniel: Yes, we could do this and it wouldn't really take that long either
Neerja: Leave out group 1 when you start with 7 hidden letters
KR: But it actually shows an interesting baseline of what happens when you just use your 7 starting letters

Final format:
Number of rounds fix to 100
Vary number of letters 0-7 (for 7 hidden letters do 200)
We will include last year's Group 1 as a player in addition to this year's 7:
8 choose 2 2-player tournaments
8 player tournament
8 player tournament, with 2 of 4 players, done enough so that everyone plays as a duplicate player once
single player tournament, with 4, 8, and 12 of each player
8 choose 2 2-player tournaments, with 5 of each player, plus a "LessStingyPlayer" who always bids rand(0,3)

If we have time afterwords, run 8 players + LessStingyPlayer

Last year:
Vary number of letters 0-6
Vary number of players 2-14
Number of rounds fix to 50

5 choose 2 2-player tournaments
5 player tournament
5 player tournament, with 2 of each player
single player tournament, with 5, 10, and 15 of each player
5 choose 2 2-player tournaments, with 5 of each player, plus a "StingyPlayer" who always bids 0

I will make the results available only as a SQL database - no more text output. I'll send out a SQLite database, probably (might make a public MySQL instead)


We did a run with 7 letters...
What went well/wrong?
Daniel: We are bidding high on letters when we are close
Lauren: We didn't do so great - we were too conservative, which is a big problem on 7 hidden letters. When we get 7 letters for free, we should make sure that we're not that conservative

KR: What makes a successful player?
EPK: It's important to adjust your bidding range based on what everyone else is doing
Monica: It should, however, be largely based upon how you value the letters.
Elba: You need to STILL value the letters carefully on a low/medium/high range, but you must define what low/medium/high is from the market
Neha: Never bid on anything that doesn't help you (ie, don't try to bid to hurt people)

KR: How many of your players will really try to harm opponents in a small game?
(4-5 people raised their hands)

Blake: Being able to adapt to the game will be important.
Yang Jian: Same - having multiple strategies

KR: How many of you have multiple strategies that you switch between?
(about half)
Zach: Our player has 3 strategies, which it switches based upon how many letters we have. We switched based upon an early/middle/late game strategy
Dfed: We are focusing on getting one main strategy to work, otherwise things become much too complex
Neerja: Concentrate on the main strategy. Work on fine tuning various parameters
KR: Yes, for sure - work hard on figuring out how to set each parameter
David: Budget control in 2 player game - make sure that you always have points to rebound as necessary
Daniel: The bonus is very disproportionate, (jumping from 11-12 points to 60+), so it's sooo important to make sure you get a 7 letter word. Spending 30 points for a 6 letter is not OK, but 50 points for 7 letters is good. It's really, really hard to figure out all of the specific corner cases and account for them.

KR: What did you use from last year?
EPK: A priori algorithm
KR: Did anyone rely on last year's strategies, with or without using the code?
Zach: The early/middle/late idea
KR: How essential was it to use last year's players for background?
Eva: We were able to bootrstrap our own discussions
Dan: We already knew a lot of the issues
Zach: It was helpful to be able to play against last year's players as we were developing
Archana: I was in some ways constrained to thinking in previous ways
Dan: It was interesting to read from past group's mistakes

No comments:

Post a Comment