Lot of cool retro video games have been long forgotten. Space Invaders, Air Raid, and Aliens are all games that us '80s kids have played. We wanted to make something that play old school video games, and beat the living daylights out of them.
We wanted to combine our love of video games with our insatiable desire to own a Tesla Model X, and decided to create an AI agent to balance a pole, on a cart.
Self-driving cars are the future, but for a car to drive itself, it has to first learn to drive in a straight line by keeping the steering wheel straight. For this hackathon we decided to make a POC reinforcement agent that can learn how to control a dynamic system with many inputs, without explicitly telling it what the rules were. Just like how the pole cart game has a state (speed/position of cart, speed/angle of pole), and a goal (don't tip the pole over), the same parallel can be drawn in self-driving cars, their state being a combination of many variables (speed of car, proximity of car to other drivers, position of car, other sensors, etc.), and the goal being to not to crash.
What it does
It plays old school video games. We started with Cart Pole,
How we built it
Deep reinforcement learning using convolutional neural networks with Keras-rl.
Challenges we ran into
Challenges? What are challenges?
Accomplishments that we're proud of
What's next for Polly
Atari applications using our general case neural network algorithms.