I am interested in sample-efficient reinforcement learning. That is, decreases the number of interactions an agent needs with an environment to achieve some goal. In the Fall of 2019, I approached this by integrating interactive demonstration data into the optimized Deep Q-Networks algorithm.
The results are positive and are heavily documented through the following:
Thanks to my advisor Dr. Ron Zacharksi and my committee members for all their feedback on my work!
In the summer of 2019, I became interested in having the interactions with the environment be in a separate process. This inspired two different implementations, ZeroMQ and HTTP. Given the option, you should use the ZeroMQ implementation since it contains less communication overhead.