Related Experiment Videos
An application of Hamiltonian neurodynamics using Pontryagin's Maximum (Minimum) Principle
1Department of Computer Science, University of Wollongong, NSW, Australia.
International Journal of Neural Systems
|December 1, 1995
Summary
This study introduces a novel learning rule for artificial neural networks based on Pontryagin's Maximum Principle (PMP). This PMP-based method optimizes system weights and shows promise compared to Standard Backpropagation for complex problems.
Area of Science:
- Artificial Intelligence
- Control Theory
- Computational Neuroscience
Background:
- Classical optimal control theory provides powerful tools for system optimization.
- Artificial neural networks (ANNs) require efficient learning rules for weight optimization.
- Existing methods like Standard Backpropagation (SBP) have limitations in certain applications.
Purpose of the Study:
- To develop a new ANN learning rule using Pontryagin's Maximum Principle (PMP).
- To assess the applicability of the PMP learning rule to discrete-time, continuous-time, and feedback systems.
- To compare the performance of the PMP learning rule against SBP.
Main Methods:
- Derivation of ANN weight update equations from PMP.
- Application of derived equations to ANNs.
- Testing the PMP learning rule on the XOR problem and comparing with SBP.
Main Results:
- A novel PMP-based learning rule for ANNs was successfully derived.
- The rule is applicable to both discrete- and continuous-time systems, including feedback networks.
- Preliminary results indicate favorable performance of the PMP rule over SBP on the XOR problem.
Conclusions:
- Pontryagin's Maximum Principle offers a viable framework for developing advanced ANN learning algorithms.
- The new PMP learning rule presents a promising alternative to traditional methods like SBP.
- Further research is warranted to explore the full potential of PMP in ANN training.