Neurosmash

A plausible deep learning reinforcement agent for the Neurosmash environment.

Deep Q-Learning Agent

Two reinforcement learning agents have been implemented. A conventional deep Q-learning agent and a biologically more plausible agent using Hindsight Experience Replay. Both agents make use of the following techniques:

• Convolutional Layers
• Backpropagation
• Frame Stacking
• LP Pooling
• Huber Loss
• Frozen Target Network

Aside from this, the vanilla implementations used

• Memory Replay

The biologically plausible agent implements

• Hindsight Experience Replay

Requirements

To run the program, make sure you have installed the dependencies listed in environment.yml. We recommend creating a conda environment for every project. You can do this with the following command: conda env create --file environment.yml

Usage

To run the networks take a look at the python notebooks in the src folder.