The inverted pendulum is a classical non-linear control problem. The dynamics of an inverted pendulum forms the basis of many phenomena, such as walking, aircraft roll control and planar robot arm control. The inverted pendulum has been used as a platform for testing the efficacy of various types of controllers. However, the controller design for an inverted pendulum is difficult because of the multi-variability and inherent instability.
In this project, the NeuraBASE network model was used as an alternate approach in controlling the swing up and balancing operations of an inverted pendulum. The approach utilised NeuraBASE to train the sensor neurons obtained via a rotary encoder and motor neurons from a stepper motor which rotated the swinging arm. A new layer, the Controller Layer, used for reinforcement learning, linked the sensor neurons to the motor neurons. In the absence of a dynamic model and theoretical control methods, the NeuraBASE network model was able to successfully control the pendulum swing up and balance operations.
The results obtained showed that a single NeuraBASE was able to adaptively balance even with the introduction of dynamic changes to the system.
Inverted pendulum system learning process.
Inverted pendulum system with external disturbance.
Inverted pendulum system adaptive learning.