Learning Methods for the Unstable Chain
In the following two pictures we compare the control performance for a system
with an unstable first mode for a learned interaction structure and a global
structure. Both simulations start with the same peak intitial disturbance where
the agent in the middle of the chain has a unit displacement and the velocities
of all agents are zero.
The left picture shows the performance of the learned structure and in the right
picture we see the control dynamics of the global structure for the same system.
The power consumption for the learned structure is only about half the one for
the global structure.
We run a series of simulations that start with a global control structure
without a market and confront the system with several disturbances in a row so
the used hill-climbing learning algorithm will find a better control structure
for all kinds of disturbances in average.
We observe the power consumption of our model system during the control proccess
for slightly different interaction structures. We produced the small changes in
the interaction structure randomly.
The system learned which structure is better suited by using a simple hill-climbing
algorithm. We started with a global initial structure and the animation above
shows how the interaction structure changes while the system learns.
last changed: July 31,1997
Oliver Guenther