• A Move Generator: Given a board position, the move generator constructs a list of all possible moves. The number of possible moves is the single largest factor determining how difficult it is for a computer to play a game. In easy games such as Chess, the number is small. In difficult games such as Go, the number can be very large. Tantrix is in the middle.



• An Evaluator: Given a board position, the evaluator estimates how good the position is for the player who is to move. If the evaluator gives good estimates, then the Robot will probably play well. The quality of the evaluator is the key factor in any game playing program.



• A Lookahead Manager: which starts a given position, calls the move generator, calls the evaluator for each possible move, and decides to accept that estimate as final to recursively call itself to examine responses, or responses to responses. If the evaluator were perfect, the lookahead manager would be unnecessary (because the evaluator would already have given the highest rating to the best move).

Back to previous page

So, how does the evaluator estimate the final score?

• First, obviously, each line is considered separately. The maximum estimate for any line is the estimate for the final score.


• For each line, consider what other, nearby lines are certainly connected, or probably connected. Groups of lines that will eventually be connected are treated as one long line, modified by the probability that the connections will actually be made. The Robot estimates the probability mostly by counting the remaining tiles that fit the connection space. This is an important point: the Robot always counts tiles. You can bet that if the Robot plays a forced connection, it knows if there is still a tile to make the connection.


• Every group of lines either forms a loop or has two open ends. By default the open ends are predicted to be extended by a fair share of the tiles left containing the desired colour. The Robot doesn't care exactly how or where those tiles will be played, it just guesses that they will be played somehow. This strategy is why the Robot prefers to extend a line rather than try for a small loop, even though the small loop would instantly double the current score. This fair share estimate is greatly reduced if the side is controlled, or likely to become controlled.
  
  
• Goodbot adds a few considerations to the basic evaluation, each of which contributes a small amount to the overall improvement over earlier robots.



• Be aware of the presence of lookalikes for forced spaces, which affect the likelihood that a connection will be made.
• Be aware when the last few tiles will certainly end up in your own rack, so probability of controlling them is 100%
• Add a small penalty for small loops in the opening.
• Delay playing unfavourable moves as long as possible.
• Consider forcing multiple tiles to be beneficial before endgame, bad after endgame.
  

Back to previous page