Keng Wah Loon - Deep Reinforcement Learning Workshop

SoCreate makes a tradition out of supporting tech in San Luis Obispo. One reason is that we’re lifelong learners, and there is so much to learn from each other, and we also want to see the tech industry grow in SLO because that only helps us develop a more talented team! So, when two experts in their field stepped up to teach a PyData SLO workshop on deep reinforcement learning, we stepped up to provide the pizza! We also hosted this March Meetup here at our SoCreate offices. SoCreate Software Engineer John Jenson attended and reported back.

Laura Graesser from Google Brain and Keng Wah Loon from Machine Zone visited SLO from the Bay Area to lead the workshop on Deep RL, which is a general framework that can be applied to solve sequential decision-based optimization problems through trial and error. You may have seen Deep RL in action when it made headlines in 2015 for AlphaGo, a computer program that beat world-famous ‘Go’ player and European Champion Mr. Fan Hui. AlphaGo went on to a 4-1 victory in Seoul, South Korea while 200 million people watched on, learning from the computer program’s inventive winning moves.

In the PyData SLO workshop, Laura and Keng covered the basics of RL, definitions of the functions that make up an RL system, and the different categories of Deep RL algorithms. “We covered two of the most important algorithms and saw these algorithms applied to basic test environments in SLM-Lab,” John said. SLM-Lab, authored by Laura and Keng, is a Modular Deep Reinforcement Learning framework in PyTorch. “The presenters gave a detailed explanation of the algorithms and functions, while also explaining how they worked on a more intuitive level.”

John said he was most impressed by Laura and Keng’s demo of a pre-trained algorithm (pre-trained because it can take a full day to train on a laptop) that had learned how to play Atari Breakout. “In a Reinforcement Learning system, an agent takes actions to affect its environment, one step at a time. Certain actions will reward the agent,” John explained, adding that teaching the system to play games is a common application of RL. The system doesn’t know the rules of the game, all it needs is to see the screen, a list of which actions it can take, and a definition for the reward. “By simply playing the game over and over, thousands or millions of times, the system learns how to maximize its reward, becoming very skilled.”

PyData SLO is an educational organization focused on data science with Python, bringing together Python users from the academic community and industry in SLO to present, share, and workshop ideas in areas including but not limited to: data science, computational sciences, machine learning, and distributed computing. We’re thrilled to see PyData bring in experts like Laura Graisser and Keng Wah Loon for workshops like these, right here at SoCreate!

Want to learn about upcoming meetup topics and locations? Click here.