Safe Action Control With a Mediating Safety Module
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Determining safe actions for technical systems that contribute to desired behaviors while ensuring safety goals is a non-trivial problem, as safe actions may not always align with achieving the desired behavior, potentially being detrimental to the system's objectives.
Innovation Solution
A computer-implemented method for training a machine learning system that determines a control signal for technical systems by obtaining a state signal, determining a distribution of potentially unsafe actions, sampling from this distribution, and using a safety module to transform potentially unsafe actions into safe ones based on a set of safe actions, with the loss value characterizing a reward for achieving safe actions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a control system determines actions to achieve desired behavior, then the system's productivity and goal achievement improve, but the safety of actions may deteriorate as safe actions may not always align with desired behavior
Solution Approach 1:
The patent introduces a safety module as an intermediary component between the policy module and the environment. This safety module receives potentially unsafe actions from the policy module, evaluates them against safety specifications, and transforms them into safe actions. The safety module acts as a mediator that reconciles the conflict between achieving desired behavior and ensuring safety by translating actions from the policy module into safe executable actions without requiring redesign of the entire control system.
2Reliability
If a safety module transforms potentially unsafe actions into safe actions, then action safety improves, but the complexity of the control system increases
Solution Approach 1:
The control system is segmented into distinct functional modules: a policy module that determines actions to achieve desired behavior, and a safety module that ensures safety of executed actions. This segmentation allows each module to specialize in its function - the policy module focuses on goal achievement while the safety module focuses on safety verification and transformation. The modular architecture reduces overall system complexity by creating independent, well-defined components with clear interfaces.
Solution Approach 2:
The safety module serves as an intermediary that adds safety functionality without requiring fundamental changes to the policy module. By positioning the safety module as a separate layer that processes actions from the policy module, the system avoids the complexity of integrating safety checks throughout the entire control architecture. The intermediary safety module handles all safety-related transformations in one centralized location.
3Adaptability or versatility
If the policy module determines actions based on state signals, then the system's adaptability to environment improves, but the likelihood of determining unsafe actions increases
Solution Approach 1:
The safety module acts as an intermediary layer between the adaptive policy module and the environment. The policy module maintains its adaptability by freely determining actions based on state signals and desired behavior, while the safety module intercepts these actions, verifies them against safety specifications, and transforms unsafe actions into safe alternatives. This intermediary approach preserves the policy module's environmental adaptability while adding safety guarantees.
Data Source
AI summary
A computer-implemented method for training a machine learning system. The machine learning system is configured to determine a control signal characterizing an action to be executed by a technical system. The method includes obtaining a safe action to be executed by the technical system including: obtaining a state signal; determining, by a parametrized policy module of the machine learning system, a distribution of potentially unsafe actions that could be executed by the technical system; sampling a potentially unsafe action from the distribution; obtaining, by a safety module of the machine learning system, the safe action. The method further includes determining a loss value based on the state signal and the safe action; and training the machine learning system by updating parameters of the policy module according to a gradient of the loss value with respect to the parameters.