Skip to content

Latest commit

 

History

History
 
 

cars

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Screen-capture of cars

Advanced car demo

This is a car demo, that shows the capabilities of neurojs. It shows it successfully can converge to a control policy, that not only can avoid objects and obstacles, but can avoid other cars.

Through a series of sensors, cars can perceive their environment and can then respond to it.

The cars are controlled via an continuous action space, which is learned via an actor-critic model, which in this case is only 2-dimensional. One control for throttle (-1 brake, +1 forward thrust) and one for steering (-1 left, +1 right).

Exploration is done by dropout, which shows very promising results. It drops-out the last layer of the actor.

The replay buffer prioritises experiences with higher error rates.

More info on the index page.

Features

  • Save/load a car brain as/from binary data
  • Click & drag map generation
  • Advanced machine learning (see neurojs features)
  • Robust physics engine
  • Fast WebGL renderer

Future

  • Path tracking (each car pulls a line behind, little bit like TRON)
  • Tablet support
  • Save/load a map as/from binary data
  • Performance & cognitive analysis (with visual focus)