DeepTraffic is a gamified simulation of typical highway traffic. The task is to build a neural agent – more specifically design and train a neural network that performs well on high traffic roads. The neural network gets to control one of the cars (displayed in red) and has to learn how to navigate efficiently to go as fast as possible. The car already comes with a safety system, so you don’t have to worry about the basic task of driving – the net only has to tell the car if it should accelerate/slow down or change lanes, and it will do so if that is possible without crashing into other cars.
The aim is to get as high speed as 80 mph where the bes submission acheived 73.73mph. Whereas the maximum speed at time of submission at leaderboard was 76.26mph.
The three top model acheived highest speed 64 mph, 70mph and 73mph.
The following are the key parameters that lead to 73mph:
- Increase the input of agents as
- lanesSide = 3;
- patchesAhead = 50;
- patchesBehind = 10;
- Use single layer with 136 neurons followed by relu activation function
- Hyper parameter are optimized as follows:
- learning_rate: 0.0001,
- momentum: 0.6,
- batch_size: 128,
- l2_decay: 0.01
- And finally the optional parameters are set to:
- opt.experience_size = 100000;
- opt.start_learn_threshold = 50000;
- opt.gamma = 0.98;
- opt.learning_steps_total = 500000;
- opt.learning_steps_burnin = 7500;
- opt.epsilon_min = 0.65;