Implementation of "Learning to Navigate in Complex Environments"

Tensorflow implementation of Learning to Navigate in Complex Environments.

Most of the codes are borrowed from Simple Reinforcement Learning with Tensorflow Part 8: Asynchronus Advantage Actor-Critic (A3C) and Maze Navigation using Reinforcement Learning.

Requirement

• Tensorflow
• Deepmind Lab
• Scipy

How to build

1. Clone the repo https://github.com/superMDguy/lab, an UNOFFICIAL version of Deepmind Lab which supports Python 3.
2. Follow the instructions (https://github.com/superMDguy/lab/blob/master/docs/build.md#step-by-step-instructions-for-debian-or-ubuntu-using-python-3-experimental) on the Github. (Recommended using bazel 0.5.2, newer version may fail to build)
3. Clone this repo under lab/python3/.
4. Modify BUILD under lab/ folder. Add bazel instructions at the end of file

   py_binary(
   name = "nav_a3c_train",
   srcs = ["python3/nav_a3c/train.py"],
   data = [":deepmind_lab_py3.so"],
   main = "python3/nav_a3c/train.py",
   default_python_version = "PY3",
   srcs_version = "PY3",
   )

5. Train the model

   $ bazel run :nav_a3c_train --define headless=glx

   If you want to show the game during training

   $ bazel run :nav_a3c_train --define headless=false

Result