Trajectory-Centric Robust Optimization for Reinforcement Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current Reinforcement Learning techniques face challenges in trajectory-centric control for non-linear systems due to time-varying controllers and noise, leading to instability and inefficiency, as they often perform trajectory design and controller synthesis separately, failing to exploit uncertainty in the system model for robust performance.

Innovation Solution

A computer-implemented method for optimizing a local time-invariant control policy using uncertain statistical models, which learns a stochastic predictive model, estimates uncertainties, and computes a trajectory-centric control policy by solving a non-linear optimization program to minimize costs and stabilize the system, thereby simultaneously designing a robust policy and trajectory.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If trajectory design and controller synthesis are performed separately, then the computational complexity is reduced, but the robustness and optimality of the control system deteriorates

Engineering Contradiction:
Improvecomputational complexityVSAvoidrobustness
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent merges trajectory design and controller synthesis into a unified simultaneous optimization framework. The joint optimization problem formulates both the nominal trajectory and stabilizing controller as coupled variables, allowing the system to exploit model uncertainty structure while achieving robust optimal control. This integration resolves the contradiction by demonstrating that the computational burden is justified by the significant improvement in robustness and optimality that cannot be achieved through separate design steps.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent transforms the control problem by changing parameters from deterministic to stochastic, incorporating model uncertainty explicitly through learned statistical models. By formulating the optimization in terms of nominal trajectories and stabilizing controllers that account for uncertainty distributions, the system achieves robust performance while maintaining computational tractability through gradient-based methods.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If a local state-dependent policy is designed to stabilize the trajectory, then the system robustness improves, but the data efficiency and computational cost worsens

Engineering Contradiction:
Improvesystem robustnessVSAvoiddata efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent segments the control policy into two distinct components: a nominal trajectory generator and a local stabilizing controller. This segmentation allows each component to be optimized for its specific function - the nominal trajectory for optimality and the stabilizing controller for robustness. By dividing the policy in this way, the system achieves robust performance without requiring excessively complex data collection, as each segment can be trained and optimized independently with focused data requirements.

Inventive Principle:
Principle #1Segmentation

3Adaptability or versatility

If the controller is made time-varying to handle non-linear systems, then the adaptability improves, but the stability and ease of operation deteriorates

Engineering Contradiction:
ImproveadaptabilityVSAvoidstability
Core Design Contradiction:
Adaptability or versatilityVSStability of the object's composition

Solution Approach 1:

The patent applies dynamics by making the nominal trajectory time-varying to capture the evolving nature of non-linear systems, while the stabilizing controller operates in the deviation space to provide consistent stabilization. This dynamic approach allows the system to adapt to changing system characteristics through the time-varying nominal trajectory while maintaining stability through the stabilizing controller that actively corrects deviations from the nominal path.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3924884B1System and method for robust optimization for trajectory-centric model-based reinforcement learning
Publication Date: 2023.08.30 MITSUBISHI ELECTRIC CORP
  • EP3924884B1 patent drawingFigure 1A
  • EP3924884B1 patent drawingFigure 1B
  • EP3924884B1 patent drawingFigure 2

AI summary

A controller for optimizing a local control policy of a system for trajectory-centric reinforcement learning is provided. The controller includes performing steps of learning a stochastic predictive model for the system using a set of data collected during trial and error experiments performed using an initial random control policy, estimating mean prediction and uncertainty associated, determining a local set of deviations of the system using the learned stochastic system model, from a nominal system state upon use of a control input at a current time-step, determining a system state with a worst-case deviation, determining a gradient of the robustness constraint, providing and solving a robust policy optimization problem using non-linear programming to obtain system trajectory and stabilizing local policy simultaneously, updating the control data according to the solved optimization problem, and output the updated control data via the interface.