• 

  StreetView

• 

  SideView

• 

  TopView

SDHacks 2017 Submission

This is a project designed for the practical use made at SDHacks 2017. We wanted to do something with machine learning that would help improve everyone's lives. What better way to do it, than to solve the daily commute?

TrafficHeaven

This is a Multi-Layer Perceptron (MLP) with a genetic algorithm to decide favorable MLP weights. The cost function we are optimizing is a number of cars in each lane of an intersection and their wait times. We then get 8 outputs and prioritize the ones with the lowest cost function, attempting to mutate and generating the next generation.

NOTE: Currently not fully functional. Neural Network needs significant tuning. But through generations, you can notice a beneficial change in simulation traffic congestion

Created by: Graham Thomas, Michael Bridges, Luke Thomas

Time spent: 24 hours spent in total

Video Walkthrough

Here's a walkthrough of implemented user stories:

GIF created with LiceCap.

Future additions

• Tune the AI
• Better Mutations
• Multiple simulations running in parallel
• Procedural generation of intersection layout

SKILLS USED

• Blender
• Photoshop
• C#
• UI
• Unity Game Engine
• Genetic Algorithm
• Neural Network (MLP)

Built With

• blender
• c#
• unity