Sjlver / icfp-2015

Team int4_t's hexagonal Tetris player. Submission to ICFP 2015

ICFP 2015, Team int4_t

int4_t: The team of four interns. Saving the world, one hexagonal cell at a time.

There is a blog post about this code, where you can see a screencast of the AI in action.

Usage

./play_icfp2015 -f qualifying_problems/problem6.json -t 150 -v

What might interest you about this code

Generating a solution is a two-step process:

1. Moves are generated using a Monte Carlo Tree Search algorithm (SamplingAI.scala).
2. Moves are translated into letters, while attempting to include as many phrases of power as possible, using a greedy algorithm (PowerPhraseEncoder.scala).

Some nice tidbits

The time taken by the program is controlled by adjusting the number of playouts (i.e., board evaluations) done in the Monte Carlo Tree Search (MCTS). The AI measures how many playouts it can do in a given time, and estimates how many moves remain. From this, it computes the number of playouts per move.

This approach is particularly elegant because, as a side effect of doing random playouts, it is very easy to estimate the number of remaining moves.

Our playouts are very light. They are almost purely random. This is probably the area where our AI could most easily be improved.

Setup Instructions for Developers

ArtificialIntelligence:

• Install Scala IDE from http://scala-ide.org. The code needs version 2.11.7.
• Install the ScalaTest Plugin (Help > Install new Software > Select the Scala IDE update site > Scala IDE plugins > ScalaTest
• Install sbt (brew install sbt, or similar)
• Generate an Eclipse project file with sbt eclipse

Submitting solutions

for i in qualifying_problems/problem_*.json; do \
    ./submitter/submitter.sh $i 6cb4e56 -t 150 -v \
    -p 'Ei!' -p 'Ia! Ia!' -p 'Chthulu' -p 'Bigboote'; \
done