Project1 - Navigation (Udacity DRLND)

This is a project that trains a agent that can navigate in the collect bananas task. (details)

Requirements

1. Download the environment from one of the links below. You need only select the environment that matches your operating system:

   • Linux: click here
   • Mac OSX: click here
   • Windows (32-bit): click here
   • Windows (64-bit): click here

2. Place the file in the p1_navigation/ folder, and unzip (or decompress) the file.

3. Packages

   • torch==1.4.0
   • unityagents==0.4.0
   • numpy==1.18.1

Getting Started

• navigation.ipynb - training code

1. We implement a Double-DQN algorithm which selects action and evaluates action using seperate networks. More details can be seen in model.py file.
2. The network architecture is a 4 layers fully-connected network along with the ReLU activation function, the number of output units are 256, 128, 32, 4 respectively.
3. Hyperparameters. The code is similar to the DQN excersise solution, except for the hyperparameter TAU, we set it from 1e-3 to 5e-3 for updating the target network much faster, because this Pick-Banana task need less episodes to convergence (up to 2000).

• report.ipynb - Provide a description of the implementation and plot average reward curve that using the pretrained model weights.

• model.py - network and agent using Double-DQN algorithm.