Are there any high-ranking ML bots?


#1

Using a fairly simple greedy strategy, my 2nd bot (had some issues compiling it so I'm at 4) has attained a rating of ~32. Having just watched a few games against tscommander-ml where my bot stomps on it, I'm curious if anyone in the top 100 or so is using an ML bot or whether they've all watched lots of games like I did and handcrafted their bots on that basis. The defector and rabbit strategies that the big guys use seem to support this.


#2

You can see top ML bots in the ML league here: https://halite.io/league-board?id=3&leaguename=Top%2025%20Machine%20Learning&language=ML&limit=25

Looks like there's one ranked 66 so far!


#3

Interesting, I wrote a fairly simply greedy client that got into the top 25% or so of active bots without much effort (rank 345 or so). What gets interesting is watching why it loses when it does. A lot depends on the fate of the first few planets. And I noticed that despite trying to avoid obstacles, I frequently lose 2 of my initial ships from collisions. I am pretty much screwed at this point. So the next step is to improve on navigation using basic bounding circle methods before I think about what kind of ML model I use. I suspected that the top bots have similar enhancements and that the programmers have hacked quirks of the game engine rather than strictly stuck to strategic and tactical ideas (rabbits and defectors for example). So

I'm rated at 32.5 or so with this. I wonder what improved collision avoidance will do for my bot.


#4

My last submission is rank 185. The wins look good as it swarms but some of the losses are sad. This bot uses the stock navigate function, but I tweaked the parameters. I think it is the 3rd highest rank ML bot currently.


#5

Interestingly, some basic implementation of separation (even with lower extend when the ships come close to their target) or moving the initial ships (due to collisions/destruction of at least two ships happening quite often) for the first rounds to different planets massively impacts the results (12XX instead of 48X).

This is not about tactics and strategy but rather on fixing distance calculations and pathfinding.
It is sad that ML is not widely used.
It would have been certainly easier to apply when there had been a grid to move on.


#6

I am hoping some of the platinum or better players OSS some of their bots now that the competition is over so I can use them to train an AlphaHalite-like bot using my C++-based deep learning framework. My bot was (just barely) in the top 1% and I could not find any material improvements on its performance past my 1/04/2018 bot which incorporated rushing, fleeing and crude damage avoidance over solid collision detection. At its heart was little more than head for the nearest available target.

Edit: Thanks for the bots on github! @fohristiwhirl, @mlomb, and everyone else! I'll put mine up shortly.