Constraint Adaptor for Safe Reinforcement Learning Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional reinforcement learning (RL) methods fail to ensure state and input constraint satisfaction in continuous state-action spaces, leading to potential system instability and constraint violations due to the lack of consideration for physical boundaries and accurate modeling of controlled systems.

Innovation Solution

The introduction of an adaptation term in the reward function that calculates the minimum effort needed to maintain the system within a control invariant set, providing a gradual penalty for constraint violations and ensuring safety, which can be applied to both legacy and new RL controllers without requiring redesign.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If conventional reinforcement learning methods are used to learn optimal control policies, then the controller can adapt to unmodeled systems and learn from operational data, but the method cannot guarantee satisfaction of state and input constraints during the exploration stage

Engineering Contradiction:
Improveadaptability to unmodeled systemsVSAvoidconstraint satisfaction guarantee
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent introduces a control invariant set (CIS) as an intermediary concept that mediates between the RL controller's exploration needs and constraint satisfaction requirements. The CIS acts as a safe region in the state space where the system can operate without violating constraints, allowing the RL agent to explore and learn while maintaining reliability through the invariant property of the set.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent performs preliminary computation of the control invariant set before the RL learning process begins. By pre-characterizing the safe operating region based on system dynamics and constraints, the method establishes a foundation that guarantees constraint satisfaction throughout the subsequent exploration and exploitation phases, eliminating the need for conservative approximations.

Inventive Principle:
Principle #10Preliminary action

2Productivity

If the RL controller explores with boundary-valued control inputs to better learn the system, then the learning efficiency improves, but the risk of directing the system state outside the specified state constraint set increases

Engineering Contradiction:
Improvelearning efficiencyVSAvoidconstraint violation risk
Core Design Contradiction:
ProductivityVSObject-affected harmful factors

Solution Approach 1:

The patent implements feedback through the control invariant set characterization, which provides real-time information about safe operating regions. The CIS feedback mechanism allows the RL controller to adjust its exploration strategy dynamically, directing actions toward boundary regions that maximize learning while the invariant set properties ensure constraints are never violated, thus resolving the contradiction between learning efficiency and safety.

Inventive Principle:
Principle #23Feedback

3Reliability

If a complete and accurate model of the controlled system is obtained to guarantee constraint satisfaction, then the reliability improves, but the model construction requires large quantities of data and complex computation

Engineering Contradiction:
Improveconstraint satisfaction guaranteeVSAvoidmodel construction complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent extracts only the essential information needed for constraint satisfaction by computing the control invariant set directly from the system dynamics and constraints, rather than constructing a complete and accurate system model. This extraction approach obtains the minimal necessary model (the CIS) that guarantees safety, eliminating the need for large-scale system identification while maintaining reliability.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3824356B1Constraint adaptor for reinforcement learning control
Publication Date: 2022.06.08 MITSUBISHI ELECTRIC CORP
  • EP3824356B1 patent drawingFigure 1A
  • EP3824356B1 patent drawingFigure 1B
  • EP3824356B1 patent drawingFigure 1C

AI summary

A system for controlling an operation of a machine subject to state constraints in continuous state space of the machine and subject to control input constraints in continuous control input space of the machine is provided. The apparatus includes an input interface to accept data indicative of a state of the machine, a memory configured to store an optimization problem for computing the safety margin of a state and action pair satisfying the state constraints and a control policy mapping the state of the machine within a control invariant set (CIS) to a control input satisfying the control input constraints, and a processor configured to iteratively perform a reinforcement learning (RL) algorithm to jointly control the machine and update the control policy.