Presentation is loading. Please wait.

Presentation is loading. Please wait.

Xi Breakthrough Player CSE 486 B Fall 2012 Miami University.

Similar presentations


Presentation on theme: "Xi Breakthrough Player CSE 486 B Fall 2012 Miami University."— Presentation transcript:

1 Xi Breakthrough Player CSE 486 B Fall 2012 Miami University

2 Team Mike Jacobs Jiang Nuo Reuben Smith

3 Components Search Threading Depth control Storage optimization Features Feature weight optimization Funny messages

4 Search Negamax with alpha-beta – Negamax is minimax “simplified” – State evaluation is from the perspective of the player associated with the current depth – Scores are negated as search moves up the tree

5 Threading Threading is used to split up the search workload evenly between search workers – The number of workers created is based on the number of processor cores available less one Workers use copies of state, so no concurrency issues are possible

6 Depth Control We dynamically adjust depth limit after each move targetMoveTime = gameTimeLimit / averageGameMoves – targetMoveTime is calculated at the start of each game – averageGameMoves is updated after each game

7 Depth Control, cont. If search finishes in less than targetMoveTime, then the depth limit increases If search finishes in more than targetMoveTime, then the depth limit decreases Depth control uses coefficients to affect how easily the depth limit can be changed again This sometimes allows us to search to a max observed depth of 25+, we can also potentially search as shallow as 1 deep

8 Storage Optimization Instead of creating a move object for each possible move at every node in the search tree, move data is encoded inside a short – Bits 0-2 represent starting row – Bits 3-5 represent starting column – Bits 6-8 represent ending row – Bits 9-11 represent ending column – Bits 12-15 represent move flags (like capture)

9 Features 7 features implemented Notable: – Squares owned – Penetration – Spread – Conflict – Cover

10 Parametric Feature Weights Feature weights can also be passed into the object This can be used to easily set a difficulty for the engine – Ex: easy = only 1 feature on hard = multiple features optimally weighted

11 Feature Weight Optimization To optimize feature weights, we could use hill climbing – We would be hill climbing in two different places 1.Hill climb to determine optimal weight for each feature in a set of features 2.Hill climb to choose the set of features that are turned on (we will choose sets of 2, 3, or 4 features) – We would be using hill climb #1 to find a high peak for scoring a trial in hill climb #2 – Momentum would be used to terminate an instance of a climb This would take a lot of computational power and a lot of time… so we didn’t do it…

12 Funny Taunts In conclusion, we also send intimidating messages to your AI at the beginning of the game – “Hello, %s. Nice to beat you.“ – “Spill the blood of the innocent!”

13 ?


Download ppt "Xi Breakthrough Player CSE 486 B Fall 2012 Miami University."

Similar presentations


Ads by Google