Machine Learning

What is reinforcement learning? 

Maseeh College of Engineering and Computer Science, Portland State University. pdx.edu/maseeh.

What is “reinforcement learning”? 

Banafsheh Rekabdar, Assistant Professor of Computer Science, Director of the Artificial Intelligence Lab: In the Artificial Intelligence Lab, we are focusing on reinforcement learning. This is a subfield of machine learning. This is an AI that learns by trial and error. It tries different actions, sees what happens, and improves over time. 

Why does reinforcement learning matter? 

Reinforcement learning matters because many AI systems need to make decisions and choose actions, not just make predictions. This is a method for sequential decision-making problems by learning through interaction. This learning through action makes reinforcement learning so powerful. 

How does your lab apply reinforcement learning in real-world contexts? 

Shayan Jalalipour, PhD Student, Artificial Intelligence Lab: I work on robustness of these multimodal reinforcement learning agents, and that essentially means working on preventing malicious interference, and different ways we talk about, how do we make it resistant to these kinds of interferences? 

Saba Izadkhah, PhD Candidate, Artificial Intelligence Lab: A group recommendation system suggests items to a group of people, and reinforcement learning helps the system learn from feedback over time and balance those different preferences to make better recommendations for the whole group. 

Bahareh Golchin, PhD Student, Artificial Intelligence Lab: My project combines reinforcement learning with generative AI and continual learning to improve anomaly detection focusing on smarter monitoring and more useful alerting strategies. 

Maseeh College of Engineering and Computer Science, Portland State University

Thank you for watching! Learn more at pdx.edu/maseeh/power-energy

My name is John Lipor and I'm an assistant professor of Electrical and Computer Engineering. I'm receiving the Researcher of the year award for the Maseeh College of Engineering and Computer Science.

My research focuses on adaptively collecting data so we can get the highest quality information for the lowest possible cost. In one project we're working to predict geothermal favorability in the Western US, so our goal is to take all the geophysical measurements that have been collected throughout decades, bind them with a little bit of expert knowledge from the geophysics community, and use all that information to predict where one could possibly create a geothermal power plant.

Estimates say there could be as many as a thousand power plants developed in the Western US so we've maybe found a quarter of them. And our hope is that through using data science and tools for machine learning that we're developing, that we can find the other three quarters of those systems.

Another project that I'm working on involves characterizing the seabed on a really massive scale, so throughout swaths of hundreds of thousands of kilometers across the ocean. The data we have to learn from is actually ambient noise in the ocean. So we might send out an underwater glider, and this glider can stay out in the ocean for months at a time. And what it does is it listens. It uses the noise generated by surface waves or maybe other sources of noise in the ocean. And from that noise we wanna determine whether we can learn about what the bottom of the ocean looks like.

Of course, to do this, the ocean is huge. And if we wanna learn about large regions, we have to be very intelligent about where we sample. Meaning we have to be intelligent about where we send our glider to take measurements. So a big part of this project is developing algorithms that tell the glider, based on all the historical measurements it's obtained so far, you've collected this information. Where should you go next? Where can we sample next to be most efficient in terms of getting the best picture of the seabed in the least amount of time?

In a side kind of outreach project that we have going, we're trying to put together an app for monitoring trail hazards in Forest Park.

So in that case, the sensors are actually, are humans that go out and as they hike through Forest Park and examine the trails, they can through the app, they can alert managers of Forest Park as to whether the trail is washed out or whether there's a hazard in the way. And so in that way, the humans are actually acting as the sensors.

Part of what I love about this research is that we can use these amazing tools for machine learning and data science to actually benefit the environment. Essentially, I'm employing similar algorithms, in some cases, the same algorithms that Google and Amazon are using to feed you ads and try and sell things more effectively; I'm trying to use those ideas and have them teach us how we can sample the environment more effectively.

My name's John Lipor and my research aims to provide a better understanding of the environment through the use of machine learning.