Hello folks, and welcome, to this, the last installment of my blog for this term’s project. Above is what I consider to be my group’s crowning achievement for the term: the car apparently dodging a cat. We haven’t seen the car behave reactively at all, and I frankly didn’t think we’d see anything like this this term. This has been an unusual project, and in many ways has been a research project. It became apparent late in the term that we just didn’t have enough runway to train the car appropriately using a Deep-Q network (DQN). I averaged about 15k timesteps per day, and the DQN defaults set in the AirSim project are set to not even start learning until 300k timesteps, and train for epochs of 500k timesteps. By my calculations, one epoch would take about a month and a half, and these are 10 week terms. To boot, there was enough research and setup to do before we could even start training, that we only had a month at best to train. Couple this with the mistakes I made in training, having to restart training several times, losing the model data several times, and, well, it’s really not surprising that our car is only performing this well.

Still, I still count this project as a success. It wasn’t my first time working in a group at OSU, but it was my first time managing a long-running project with a group of more than two people. The only other project that came close was in Databases, but that was a lot easier to handle, with two people. The logistics of making sure that everyone had something to do, but not too much to do, were challenging. It was cool, though – most of my group is employed in industry already, so it was nice having standards around commitment, accountability, and the Agile ceremonies. So, I learned a lot about working in a group. I also was exposed to the huge and academically deep discipline of machine learning and artificial intelligence in general. So many of the world’s brightest minds are working on these problems, and it was such a pleasure to get to experience part of what is being worked on right now. I only scratched the surface of understanding how neural networks work, but we still managed to leverage one to accomplish a task. In my view, this makes us no less than wizards.

It’s hard not to pause and reflect here. This is my last class, and in less than two weeks I’ll have completed the degree program. Raising a family, working a challenging job, and earning an engineering degree at the same time has been challenging in a way that words can never do justice. It’s been a team effort – I’ve gotten so much support from my immediate and extended family, dear friends, classmates, some really top-notch instructors, I’ve been saved by diligent TAs, tutors, the Success Coaches, and I owe my current position to the OSU Corporate Relations department. And I couldn’t have done it without my trusty sidekick, my feline friend Ursula. What a list! As for what’s next: some time with my family, who haven’t seen much of me in the last 3 years. And time and tide march onwards…

Thanks for reading, and good luck to you in whatever you choose to do. Happy coding!