If neural networks don’t predict recommended actions, then how do self driving cars make their predictions?
Neural Networks do predict recommended actions. But that comes under separate subdomain of ML known as Reinforcement Learning. There is a field Deep Reinforcement Learning where an agent learns about an environment/game using NN’s(search for DQN’s). Ex-AlphaGo.
thanks, so which is used in self driving cars?
Imagine a little car whose objective is to drive following a black line; you can equip the car with a sensor and determine the track’s colour, then the agent can use that information to align the car to the black line based on that information. Similarly, a real car can equip with sensors and machine learning models to digest and interpret that information. The agent has more than the raw output information from sensors; it can also have predictions from a model that used sensor information to recognize objects like a stop sign. The agent can take actions using these predictions, like stop the car. In the case of self-driving cars, you have an agent that considers many information processed (from many models) and used that information to take actions within a given environment. Therefore, it’s not the model that recommend predictions, is an agent that used sensorial information from model to recommend a decision: if (this_model_detect_stop_sign): suggest to stop the car / or stop the car if you are in charge.
Thank you very much for your reply. So does this part in self driving cars get hard coded. I just want to understand the mechanics
Omar,
No, I’m afraid hard coding that would be a disaster. (Like hard coding “this is a cat”, but with far higher error costs.)
The deep net learns to associate inputs with actions by watching or doing. I haven’t built one yet, but these look like good starters:
- Self-driving cars in PyTorch - blogger reworks a Udacity exercise
- KDnuggets 5 links to get started w/self-driving cars
One of my co-workers used to use reinforcement learning to play Doom, Starcraft, and similar things.
Thanks man! I really appreciate your help! 