Deep RL Bootcamp Lecture 10B Inverse Reinforcement Learning

 

 

 

 

 

 

 

 

 

 

 

 

 

 

intrinsic ambiguity: move toward purple triangle? move away from red triangle? move along grey arrow? or the combine of them?

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

the right part of the right top equation should be divided by T

 

 

 

 

 

 

 the original paper used linear function, but equations here aren't limited to linear function.

 

 

 

 

 

 

 

 

 

 

 

 

Ho & Ermon's 2016 paper has similar high level structure

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

robot execuated tasks under the help of human

 

 

 

standard cost/reward function in the algorithm like policy search

cost function is the negetive of the reward function 

this video will fail. this isn't to say that we can't that handle of reward function for this task, but require a lot of  trials and errors. a lot more difficult than demonstrating how to do that by hand. 

 it's quite aggressive, because it's trying to get to the goal position as quickly as possible. 

 

 

 

 

this is the result, if you run  cross entropy 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

here is the result running relative entropy in RL. Again, it didn't get a good sense of how to successfully perform the taskj.

 

 

 

 

 

 

 

 

 

orange: behavior cloning

blue: GAN ...???

 

 

 

 

 

 

 

 

 

 

Linitations, NO.2: most data used here is low-dimensional feature spaces

 NO.2: first person point of view, not what the robot sees

 

 

 

 

 

 

 

 

 

 

 

 

 

转载于:https://www.cnblogs.com/ecoflex/p/8990309.html