Captain Cwack: A Reinforcement Learning Agent

• 

  Gameplay

• 

  Average Reward

Inspiration

I have been working on Reinforcment Learning problems for a little over a year now and wanted to find a cool project that I could demo at GH.

What it does

An RL Agent controls the jetpack for Captain Cwack guiding him through obstacles on his way to deliver the results of the grizzhacks competition. Captain Cwack is also playable by humans (and there is a leaderboard)

How we built it

Captain Cwack's base environment is built within pygame, this environment is wrapped within a gymnasium environment, and then a Proximal Policy Optimizer (PPO) is used to find the optimal strategy for him.

Challenges we ran into

So far the largest challenge is getting the agent to perform at the level of humans, currently humans can average between 700-1200 points per round while the agent is at 300, there are many reasons this could be the case, but additional testing and experimentation will be needed.

Accomplishments that we're proud of

It works!

What we learned

Taking a systems approach to building complex environments will keep the code clean. Having a very detailed plan before attempting implementation saves many headaches.

What's next for Captain Cwack: A Reinforcement Learning Agent

He's gotta deliver the results of this Qwackathon! As for me, I continue to pursue studying reinforcement learning problems in my free time outside of work. This should be increasing soon as I will be graduating in a couple of weeks, so I am pretty excited to dive into some more advanced libraries and more complex environments!

Built With

• gymnasium
• pygame
• python
• stablebaselines3