Starring... Felienne Hermans
Felienne is assistant professor at Delft University of Technology, where she makes programming for non-programmers more awesome. She built an IDE for spreadsheets in the form of smell detection, refactoring and unit testing tools for Excel, and she has researched code smells and clone detection in the Scratch programming language for kids.
She is also one of the founders of Joy of Coding, a one day developer conference in Rotterdam, celebrating the love for programming. When not building programs that play board games or card games, she is probably dancing Lindy Hop or killing creeps in a tower defense game.
So, AI bots have long mastered computer chess, and recently a neural net beat one of the world’s best Go players. What fun is there still to be had in machine learning for games? Well… What about your granny’s favorite afternoon entertainment: Bridge?!
Bridge is a card game with two distinct phases: bidding and playing. For this talk, Felienne will focus mainly on the bidding part, as that is most challenging. In the bidding phase, both pairs of players bid to reach ‘the contract’: the number of tricks they want to make, and with which trump color. Given the limited bandwidth of communication (players can only communicate with bids) the challenge is to get at the best bid. She made a DSL in F# to describe the bidding rules that her bot will bid with, and she will talk about its design and the choices she made.
For bidding, there are a number of standard systems, but of course she wanted to go a bid further*. Reaching the optimal bid is very important: Failure to make the contract results in a penalty, but not reaching a possible contract does too. Therefore, she used genetic programming to combine different existing bidding strategies to reach the perfect scheme.