As stated in the title, I'm wondering if it would be possible to "outperform" the master in the apprenticeship learning. I'm aware that the question might be not clear enough; but hopefully, someone might have done something on it before.

More precisely, I'm actually asking a clear way to define the target of learning: If the master's behavior is defined as standards, then of course the answer should be NO.

However, consider a real example: in the case of autonomous vehicles, if we design the algorithm to mimic a human master, then would it be possible for the algorithm to outperform the master (consider ideal case, and neglect physical condition, e.g. tired, un-focused...), especially in new situations, if we well-define a new reasonable standard?


Apprenticeship learning is equal to record a trajectory in an abstract state space and reproduce it later. For example, if a human operator is driving a remote controlled car in a circle, this pattern is the goal behavior. The autonomous algorithm take the goal pattern as an input trajectory and reproduce it later. That means 24/7 and with a higher speed of the car. Outperforming the master is not the exception, it is the reason why Apprenticeship learning is realized in robotics.

To the second part of the question,

quote: “If the master's behavior is defined as standards, then of course the answer should be NO.”

Perhaps the idea is to implement a safety robot control system which do not harm the environment. For example, a fast driving car in a circle will produce a lot of noise. So the idea is to train the system to fulfill industrial criteria. According to the definition, the system is not allowed to drive faster than the human operator.

There is some kind of universal standard, which can be cited as the law for evolution in robotics. Any robotics system which is aggressive to it's environment will not be tolerated. Only social robots which are cute will survive. Any system which violates the standards will be deactivated after 5 minutes. In contrast, a system which is more adorable than expected by the designer has the chance to life very long. Because the environment asks for amusement.

| improve this answer | |
  • $\begingroup$ Thx Manuel. I guess my question should be claimed this way: given one or several trajectories for the robot to learn, say, all of which are produced by the same human athlete who can, according to a designed reward-rule, always get scores. Then let's consider a new task for both the athlete and the robot, while this time we score them according to the rule we used before on grading the human athlete, would you expect a high score obtained by the robot? Or is it possible to achieve so (without a direct reinforcement learning)? $\endgroup$ – Kite.Y Oct 18 '18 at 12:34
  • $\begingroup$ Transfer learning of a human athlete is much more complicated than only control a car or a helicopter. You will need knowledge about kinesiology which is stored in grammars and Central pattern generators for the trajectories. It is also called apprenticeship learning too, because it has a capture- and a reproducing phase. $\endgroup$ – Manuel Rodriguez Oct 18 '18 at 13:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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