When people start to figure the self driving cars will replace some vehicles on the road in the very near future, it somehow means the learner in the robot software could reach a very low empirical error to ensue the safety of passengers. The great deal of driving data collected and the deep reinforcement learning algorithms contribute mostly to the success of the self driving practice as we see it.

When it comes to many cases of our own practice, we yet cannot build up a predictive model to such a high accuracy to bring the huge difference to our own business.

My question is, when I saw the machine learning techniques can make revolution on the car driving due to its successful algorithms to implement behavioral cloning tasks well, what is the foremost reason for this success, the big training data set or the deep reinforcement learning techniques or any particular reason in self driving problems?

Furthermore, can this success be copied in most of the practical problems we solve in machine learning? In other words, if today's machine learning techniques can help a car drive by itself, how can I reach the same success when developing my own predictive models to boost my business? If we cannot, what are the constraints, not enough data in the particular case or not a smart model yet? The reason the self driving thrives can help answer this question.


1 Answer 1


I'll give a high level answer.

Our Bias on Effort

Although self-driving technology is now on the verge of becoming mainstream, incredible effort has gone into its development. The DARPA Grand Challenge, which helped in this technology's advancement, started in 2004. Imagine the total effort that went into this technology over 13-14 years. Many labs were setup with sole aim of competing in this challenge.

So when you say that we can't replicate the success in our businesses directly, it is better to ask if the comparison is fair. In your business, have you worked on a single problem to such an extent that many bright minds have worked on it for many years? If not, the comparison is not fair (to your business!).

Self-driving cars are not overnight success due to a single algorithm.

No Single Secret Sauce

Media reporting makes us believe that there's one algorithm that rules the technology, reality is more complex. Check this review paper on the complexity of tasks. There are many modules involved (sensors, perception, planner and control module, feedback loops). Yes, probability and machine learning plays a large part in all modules, but there's no single oracle algorithm. Old workhorse algorithms like Kalman Filters also shoulder heavy workload, but wouldn't get mentioned in the deep learning hype.

Yes large data helps, reinforcement learning helps. But so do rules inferred from manual interventions and control theory and older algorithms.


Many areas from engineering (classic control theory, reinforcement learning, machine learning) form the basis of self driving cars. We can't copy over the success to other domain easily: there's no single secret sauce but tons of hard work and solid engineering over long time.

  • $\begingroup$ Thank you @hssay for your answer. When I talked about the success of self driving, I view it as a machine learning application case and I was highly impressed by the model's performance. So, based on this fact, I would like to generalize the drivers for this machine learning success: the data or the algorithm or any other factors? So, in my mind, maybe there is an answer in a format like: because point 1, point 2... the self driving modelling can be of success, But in our own case, there is some difference that would restrain our models to perform that perfectly. $\endgroup$
    – LUSAQX
    Jan 4, 2018 at 20:55
  • $\begingroup$ Indeed, my question is expecting a high level answer in terms of a general methodology about the successful machine learning modelling. Overall, if machine learning can help a car drives by itself, why I cannot always use machine learning to build a model to predict a business problem well enough? I don't get enough data, not having smart features or not developing an advanced model? If I deem the self driving a benchmark for machine learning performance, then I would like to know why my model fail to reach that success and what kind of efforts I can make as people do it in self driving? $\endgroup$
    – LUSAQX
    Jan 4, 2018 at 21:08
  • $\begingroup$ To my one last year experience in machine learning, I personally figured it is generally two main factors behind any machine learning practice: the data (signal) and the algorithm. But what I still don't know is the specific data characteristics and algorithm design in self driving success. So, maybe someone with professional self driving knowledge can demonstrate it a bit more and then we can take it as a reference to see why we cannot copy it in our own cases. $\endgroup$
    – LUSAQX
    Jan 4, 2018 at 21:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.