Doom Reinforcement Learning Agent

Using a DQN training on a CPU for about 5 hours, I trained an agent to play the Defend the Center scenario in the ViZDoom simulator.

Repo: rl/dtc



VelmaX

This is still an ongoing project. It started as the final project for the Deep Learning class at CMU, and I collaborated with two other students, Alex Xiao and Eric Zhu. The goal of the project was to incorporate a controllable element into a dialogue agent. This would allow someone to ask for a positive or negative response, or maybe after training on the sentiments of a single character’s lines in a show, allow the dialogue agent to make progress towards modeling a personality.

So far, we reimplemented VHRED in Tensorflow, and figured out how to stabilize and speed up learning learning. The next step is to incorporate the controllable model given here.

Our results for the first part of the project are given here.