Related Experiment Videos
Evolving neurocontrollers for balancing an inverted pendulum.
1Max-Planck-Institute for Mathematics in the Sciences, Leipzig, Germany. f.pasemann@mis.mpg.de
Network (Bristol, England)
|April 30, 1999
Summary
This study presents an evolutionary algorithm for designing recurrent neural networks as nonlinear controllers. These networks efficiently solve complex control problems like balancing an inverted pendulum, even with minimal network size and limited information.
Area of Science:
- Artificial Intelligence
- Control Systems Engineering
- Computational Neuroscience
Background:
- Recurrent neural networks (RNNs) are powerful tools for complex control tasks.
- Designing effective RNNs often requires manual tuning of architecture and parameters.
- Existing methods may involve quantization, limiting performance.
Purpose of the Study:
- To introduce a novel evolutionary algorithm for automated RNN controller generation.
- To simultaneously optimize network size, architecture, and parameters without quantization.
- To demonstrate the algorithm's effectiveness on the pole-balancing problem.
Main Methods:
- Development of a tailored evolutionary algorithm for RNN generation.
- Simultaneous optimization of network topology (size, architecture) and parameters (weights, biases).
- Application to the benchmark pole-balancing problem (inverted pendulum).
Main Results:
- Evolved RNNs successfully functioned as nonlinear controllers for the pole-balancing task.
- Networks utilizing recurrent connectivity excelled with reduced state information (angle, cart position).
- Minimal-sized evolved controllers achieved high benchmark performance.
Conclusions:
- Evolutionary algorithms can effectively generate high-performing RNN-based nonlinear controllers.
- Simultaneous optimization of architecture and parameters is a viable approach.
- The proposed method offers a powerful alternative to manual RNN design for control applications.