miércoles, 12 de marzo de 2014

Value Iteration

Value Iteration
Download

Value Iteration

WHAT IS IT?

A demonstration of the value iteration algorithm as applied to a 2D world where a robot can move North, South, East, or West.

HOW IT WORKS

We build a graph where the _node_s represent the state of the underlying MDP and the directed links represent actions that can be taken on each state. Each link has a transitions variable which holds the set of nodes that can be reached when taken that action, along with the probabilities.
The thickness of each edge/action is proportional to its current utility.
prob-action-works is the probability that the action (North, South, East, West) will actually take the robot to that square. With 1 - prob-action-works the robot will end up either at its current spot or at one of the other reachable nodes that is a distance of < 2 from the intended destination, with equal probability.
The plot shows the maximum change in utility over all nodes. As expected, this value decreases monotonically.

HOW TO USE IT

Setup and Go.

CREDITS AND REFERENCES

Jose M Vidal

CHANGES

20110514
Initial revision


Fuente: JM Vidal

No hay comentarios:

Publicar un comentario