Papers
Topics
Authors
Recent
Search
2000 character limit reached

Multi-Step Prediction of Dynamic Systems with Recurrent Neural Networks

Published 20 May 2018 in cs.NE and cs.RO | (1806.00526v1)

Abstract: Recurrent Neural Networks (RNNs) can encode rich dynamics which makes them suitable for modeling dynamic systems. To train an RNN for multi-step prediction of dynamic systems, it is crucial to efficiently address the state initialization problem, which seeks proper values for the RNN initial states at the beginning of each prediction interval. In this work, the state initialization problem is addressed using Neural Networks (NNs) to effectively train a variety of RNNs for modeling two aerial vehicles, a helicopter and a quadrotor, from experimental data. It is shown that the RNN initialized by the NN-based initialization method outperforms the state of the art. Further, a comprehensive study of RNNs trained for multi-step prediction of the two aerial vehicles is presented. The multi-step prediction of the quadrotor is enhanced using a hybrid model which combines a simplified physics-based motion model of the vehicle with RNNs. While the maximum translational and rotational velocities in the quadrotor dataset are about 4 m/s and 3.8 rad/s, respectively, the hybrid model produces predictions, over 1.9 second, which remain within 9 cm/s and 0.12 rad/s of the measured translational and rotational velocities, with 99\% confidence on the test dataset

Citations (82)

Summary

  • The paper introduces a novel NN-based state initialization method that outperforms traditional washout techniques for multi-step prediction.
  • It evaluates MLFC and LSTM architectures, demonstrating improved accuracy and stability in modeling complex dynamic systems.
  • Hybrid models combining RNNs with simplified physics effectively enhance prediction reliability, especially for aerial vehicle dynamics.

Multi-Step Prediction of Dynamic Systems with Recurrent Neural Networks

The paper "Multi-Step Prediction of Dynamic Systems with Recurrent Neural Networks" addresses the challenges of modeling the dynamics of complex systems such as aerial vehicles using Recurrent Neural Networks (RNNs). It provides insights into enhancing multi-step prediction accuracy by introducing an effective state initialization strategy and explores the integration of black-box neural models with simplified physics-based models to further improve prediction fidelity.

Introduction to RNN-based Dynamic Modeling

The study explores the problem of multi-step predictions, crucial in applications like state estimation and control schemes. Traditional white-box models, based on first principles, often struggle with unmodeled dynamics and require precise parameter identification. In contrast, black-box models like RNNs leverage data-driven insights to account for complex nonlinearities, though they necessitate an effective state initialization approach to ensure accuracy over longer horizons.

State Initialization Problem

A key challenge in employing RNNs for dynamic prediction is the proper initialization of network states at the beginning of each prediction interval. The paper proposes a novel Neural Network (NN)-based initialization method that significantly outperforms traditional washout techniques. This NN-based approach leverages an auxiliary network to initialize RNN states effectively by training on a dataset divided into initialization and prediction segments.

Neural Network Architectures

MLFC and LSTM Network Configurations

Two primary RNN architectures are evaluated:

  1. Multi-Layer-Fully-Connected (MLFC) RNN: Utilizes sigmoid layers with skip connections to mitigate vanishing gradient issues.
  2. Long-Short-Term-Memory (LSTM): Employs gated memory cells to manage long-term dependencies, enhanced with peephole connections for better dynamic information propagation.

Hybrid Models

The work introduces a hybrid model that combines a simplified physics-based method with an RNN. This hybrid approach aims to compensate for the limitations of purely data-driven models by embedding prior physical knowledge, enhancing prediction reliability particularly in early prediction intervals.

Experimental Evaluation

The research evaluates the performance of the proposed methods on two datasets:

  • Stanford Helicopter Dataset: Highlights the improvement using NN-based initialization, noting challenges due to limited data and unaccounted factors like wind.
  • Quadrotor Dataset: Demonstrates significant gains both in prediction accuracy and stability when using the hybrid model over purely data-driven methods, particularly for longer prediction horizons.

Prediction Accuracy and Model Improvements

The paper demonstrates that:

  • NN-based Initialization: Outperforms washout methods, providing more accurate and consistent initial predictions, crucial for effective multi-step forecasting.
  • Hybrid Model Enhancements: Offers substantial improvements in prediction accuracy for aerial vehicle dynamics, especially in reducing errors in velocity and angular rate predictions.

Conclusion

The study illustrates the potential of integrating machine learning with traditional modeling techniques to improve dynamic system prediction. The findings pave the way for more effective control and estimation strategies in robotic systems, leveraging the strengths of both data-driven learning and physical modeling.

Paper to Video (Beta)

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Collections

Sign up for free to add this paper to one or more collections.