halite

Source code of bot for Halite, an online AI competition organized by two sigma.

Ranked 9/1592 in the official final. Some replays.

I first noticed Halite in Jan 2017 from a Linkedin animation of two bots combatting. After exploring the website halite.io a little bit, I find it intellectually engaging. I tried to follow some tutorials in the forum and began to submit bots, starting from the famous overkill bot.

Every weekend I spent some time to improve the bot, adding features to overkill bot little by little. The overkill bot doesn't have traffic control, so adding a map to record movements and strength can help redirect bots from overcap. To avoid overkill from opponents, it is also useful to rearrange directions of sites near the combat border.

In the last weekend of competition, I finally sat on to really think about improvements. I made a few changes in the last two days:

1. Non Aggression Pact. Through observation and test, I found that not attacking border between self and opponents has a big advantage on the opponents which don't implement this strategy. It is simply not engaging war with multiple oppenonts in early game.

2. To avoid overkill from opponents, leave a gap between first row combat sites and second row sites.

3. Increase the wait time for inner sites from 5 to 9 if the claimed area exceeds half of the whole map. These numbers (5 and 9) are purely heuristic, and may not be optimal. Early expansion need more flexibility, but if bots get into combat stage, it usually will last for a while. Bump the wait time will increase your strength a lot and win the game in the long run.

I hesitated for publishing this not well-designed code, and even thought about refactoring the code. Then I thought the time can be spent on something better and treat it as a learning experience.

Two things that I have learned as a first time AI player:

1. Spend some time on the inital design will save time in the long run. After spending some time to refactor the code to class design, I give up because it need lot of time to tune parameters. It would be better if I have started that way.

2. Fancy algorithm doesn't necessarilly beat simple and well-tailored algorithm. Have something working is better than something might be better but hard to implement.