In the control theory, a forward model describes predicted behaviors of a system. A forward model of a car physics can calculate the position of the car with x/y value if the steering wheel is put in the middle, left or right.

enter image description here

The outcome prediction differs, if the speed is low, middle or fast. A naive approach to create such action models is with ordinary differential equations which are forming a mathematical state space. A more elaborated technique is to use machine learning for generating the forward model only with data in the loop.

The literature has introduced the term "action model learning" for the regression of a forward model. Most papers have symbolic tasks in mind which results into a PDDL action model, but in theory it's possible to create numerical values as well with the technique. The problem is, that possible machine learning algorithm are endless. I've found in the literature action learning frameworks which are working with decision trees, recurrent neural networks, inductive logic programming and MAX-SAT solvers.

Are there are some examples from real life robotics challenges available (like the following problem or Robocup soccer), in which someone has used action model learning with one of the cited algorithm and has recognized if it's working or not? Which mistakes could be made using such ML techniques, or which situations the forward model can't learned (e.g. using a decision tree, because e.g. the state space is too large)?


Your Answer

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

Browse other questions tagged or ask your own question.