Halite III, AI competition by TwoSigma. The two main ingredients for my bot were linear programming (so that agents do not crash into one another) and a simple neural net governing high-level decisions (e.g. where to set a new base or which area to direct the ships to), which was trained on self-play.

This relatively simple algorithm outperformed nearly 99% other bots in the competition. Below are some of the winning games.

Small map 1 v 1

Medium map 1 v 1

Large map 1 v 1 v 1 v 1

I made a simple ✦Neural Transfer app✦ (applies the style of one image to another) via Streamlit and deployed it on Heroku for free. You can upload your own images and the network will be trained in real time!

Since the server is free, there are some constraints. For example, I could not use VGG19-net because its weights are too large, so had to switch to Alexnet. That's why the performance may not be too impressive. It also runs slow, but does the job eventually.

App Code

---

Update

Streamlit sent me a sharing invite so I deployed the app there as well. The server runs faster than Heroku and has less stringent requirements, so now the app sports VGG19!

App Code