Ready for something completely different? The Flatland Symphony is finally here!
One of the main objectives of the Flatland challenge is to find a suitable observation to solve the problems.
Meet a radically new observation builder, inspired by acoustics. Every object in the rail network emits characteristic “sounds” which propagate through the grid, recording information about their path.
Sounds eventually converge to an agent’s “ear”. Ears buffer sounds and encode them into state observations suitably formatted to be processed by conventional RL policies.
This observation builder has been carefully handcrafted with the following goals in mind:
- Try to improve on existing observations. Specifically, be at the same time more precise and more concise.
- Adapt to both small and large environments
- Be customizable, cleanly coded and well-documented. Make life easy for interested developers.
This observation builder is experimental and still a work in progress. Take a look at the explanatory notebook in Github or jump directly into the code. The code is published under a MIT license. Inspect it, try it, extend it - it’s yours to hack!
This observation builder competes for the Community Price. It has been a lot of fun to develop it and I hope that it helps future contestants push the boundaries of RL in Flatland! Cheers!