A (maybe working?) implementation of the first part of this paper: https://arxiv.org/pdf/1603.00748.pdf, tested on the OpenAI Pendulum task. Not very well documented / organized at present. Ideally I'll be able to make it robust enough to work across many tasks with minimal tuning (which may require implementing other features described in the paper). I also plan to try integrating the algorithm into some recurrent attention models (e.g. https://github.com/jlindsey15/RAM and possibly a modified version of https://github.com/jlindsey15/DRAM).
The following were/are helpful as references -- at the moment I don't think my code doesn't offer any more significant functionality than these... but more to come!
https://gist.github.com/tambetm/78227e1a15c52fbbcaeef7715dd079f0#file-pendulum-v0-md https://github.com/carpedm20/NAF-tensorflow