crowdAI is shutting down - please read our blog post for more information
ViZDoom Reinforcement Learning
By Poznan University of Technology
almost 2 years ago
Creating a Doom fighting AIs isn‘t a mainstream task. There is just a few of us so i think it would be fun to learn a bit more about each other, who the creator(s) of the agents are and what brought each of us to ViZDoom.
— About the team —
My name is Michael Krax and I am participating as a hobby without any company or University involvement. I am a product manager for digital communication living near Boston, MA
— About the agent —
I dubbed my agent CVFighter mainly because this is my first Conputer Vision project
— Why ViZDoom —
I stumbled into a few articles last year about the competition and really liked the concept. I was looking for a new hobby project so I decided to give it a try.
almost 2 years ago |
It’s crazy how close the competition is this year. I’d be really interested to see how all bots perform against a human player.
About the team - My name is Ben Bell creator of the Marvin bot in last year’s competition and it’s just me on the team. I’m a ML researcher with a focus in reinforcement learning.
About the agent - Marv2in is an upgraded version of Marvin. The main changes this year were using an LSTM and training for much longer on many more maps. I will be releasing code and my dockerfile after Round 2 is over.
Why ViZDoom - It’s such a cool game to try and learn. Not only does it require quick reactions but also long term planning and map exploration/exploitation. I definitely think there’s still a lot of things to learn from this game and would love to see more tracks next year like 1v1 on known maps and team deathmatch (even if they are informal/without prizes).
almost 2 years ago |
My name is Anssi (“Miffyli” around internet), and I am trying to do my PhD on reinforcement/machine learning and video games.
Previously I participated with agents TUHO (2016) and TURMIO (2017): They used basic template-matching enemy detection along with DQN trained to navigate around the map. TURMIO included some manual tunings of looking at different directions every now and then, but it seemed to make things worse.
This time around we (an undergrad and me) tried self-learning-ish approaches. We restricted ourselves from using filthy ape demonstrations or other exploration heuristics, but in the end we weren’t able to train anything sensible. Most of the time policies reduced to some simple actions of just rotating and shooting. Including curriculum learning (e.g. simpler maps, lower health at beginning) did help but not enough to make a difference. I guess this is doable, but requires way more training than what we were able to do. Regardless, was fun to do! Even got some ideas for future research ^^
Oh and how I found ViZDoom: Did my bachelor’s thesis about deep reinforcement learning, and my supervisor stumbled upon this competition. At the first glance it seemed too spooky and difficult for me but he pressured me to join. Thanks to that competition I learned a bunch and kept me motivated to learn more. If I am bugging you, blame Michał and Marek for doing such an interesting competition that keeps me around.