Leveraging Fractional Dynamics in Reinforcement Learning for Efficient and Robust Control

Authors

• Matthew Tubbin, Portland State University

Funding Source

Frank&Ken (Frankwell Lin and Ken Li) and PSU ECE

Advisor

Christof Teuscher

Document Type

Poster

Publication Date

2025

Abstract

Brain-inspired neuromorphic hardware offers a path to energy-efficient computing. This research explores the Fractional-Order Spiking Neural Network (FOSNN), an architecture where a tunable parameter, α, introduces a ‘memory’ of past events into the neuron’s dynamics. The goal is to determine how memory affects a model’s computational efficiency and its resilience to both environmental noise and network damage, in comparison to simpler models. Central finding: A trade-off exists between the MLP’s structural resilience and the FOSNN’s noise tolerance. FOSNN: Best for noise and real-world data due to its memory-based filtering. MLP: Best for stable conditions, showing superior resilience to network damage. α-value: Serves as a critical tuning knob for FOSNN memory and efficiency, with α = 0.5 – 0.7 providing both strong noise robustness and efficient computation.

Citation Details

Tubbin, Matthew, "Leveraging Fractional Dynamics in Reinforcement Learning for Efficient and Robust Control" (2025). Maseeh Summer Undergraduate Research Experience. 95.