代写COMP9444 Neural Networks and Deep Learning作业assignment

COMP9444 Neural Networks and Deep Learning
Session 2, 2017
Project 3 – Deep Reinforcement Learning
Due: Sunday 29 October, 23:59 pm
Marks: 15% of final assessment

Introduction
In this assignment we will implement a Deep Reinforcement Learning algorithm on some classic control tasks in the OpenAI AI-Gym Environment. Specifically, we will implement Q-Learning using a Neural Network as an approximator for the Q-function, with experience replay.
There are no constraints on the structure or complexity of network to be implemented, and marks are awarded on the basis of the learning speed and generality of your final model.
Because we are not operating on raw pixel values but already encoded state values, the training time for this assignment is relatively short, and each problem should only require approximately 15 minutes on a standard laptop PC.

The structure of this assignment is somewhat different to Assignment 1 and 2. Instead of functions to implement in a separate file, you have been given a partially completed python script.
Once you have implemented the unfinished parts of the script, you can run the code the same way you have with python3 neural_Qtrain.py. However, you can also call your main() function from another file, or invoke it from the command line if you wish to pass in specific argument values.
If running from the command line, navigate to the directory where the file is located and call the file like so:
$ python neural_Qtrain.py “CartPole-v0”
Would run the cartpole environment.
Alternatively, you can run the code interactively in ipython with e.g.
$ ipython
In [1]: import neural_Qtrain.py as qlearn
env, state_dim, action_dim, network_vars = qlearn.setup()
In [2]: qlearn.qtrain(env, state_dim, action_dim, *network_vars, render=True, num_episodes=5)
You can then run qlearn.qtrain() again with different values for “render”, “num_episodes”, etc. without reinitialising your network.