Objective:
Create a Reinforcement Learning algorithm to learn to play and complete the track, similar to the Helicopter Game.
The state is defined by the Field of View that the Helicopter is able to percieve in front of itself. In the current configuration the is a 3x5 grid, in addition to this the model also gets given the height from the floor of the grid world.
Implementation details for the Q-Learning Algorithm have been taken from the Playing Atari with Deep Reinforcement Learning.. Notable implemented features from DeepMinds work included:
- Experience Replay
- Epsilon Decay during Training
- A version of Reward Clipping/ Scaling
Info:
Descriptions added as headers for each file. Main files to run and plot models are: main.py and plot_results.py
Models:
- e-Greedy Q Learning (1)
- e-Greedy Q-Learning with e-decay (2)
- Deep Q-Network (DQN)(3)
Files:
- Train.py - Train a Model (Saves Memory and Metrics to JSON and Pickle files)
- Test.py - Test the Model on an Unseen Track and Plots Results
- Plot_results.py - Plot results from the Models generated from the Train file
File Structure:
- rf_helicoper/
- Model/ : Contains all the Scripts create tracks, generate plots and Agent model types
- Results/ : Contains all the saved memories and plots generated
- Tests/ : Started working on integration tests (WIP)
Other Information:
The different cases can be found in the Settings.py file.
- Case 1: Create task that a Reinforcement Learning Alogorithm could solve.
- Case 2: repeating the experiment in Case 1 with different gamma values
- Case 3: repeating the experiment in Case 1 with different learning rates
- Case 4: repeating the experiment in Case 1 with different policies
- Case 5: repeating the experiment in Case 1 with different state and reward functions