Navigation
Minimap of introduction diagram
Minimap of stage diagram

AMLAS outline

ML deployment argument pattern

The argument pattern relating to this stage of the AMLAS process is shown in Figure 16 below. The key elements of the argument pattern are described below.

Figure 16: Assurance Argument Pattern for ML Model Deployment

It must be demonstrated that the safety requirements allocated to the ML component are still met when the ML component is deployed to the system in which it operates. This is shown by providing two subclaims. Firstly, the ML component integration claim demonstrates that the safety requirements (that were satisfied by the ML model) are also met when the ML component is integrated into the rest of the system. Secondly, the ML component operation claim is provided to show that the safety requirements will continue to be met throughout the operation of the system.

It must be demonstrated that the safety requirements allocated to the ML component are satisfied when the component is integrated to the system. To demonstrate this, the ML component must be executed as part of the system following integration. It must be checked that the safety requirements are satisfied when the defined set of operating scenarios are executed. The operating scenarios used in the integration testing ([FF]) are provided as context for the claim. The sufficiency of the operating scenarios that are used must be justified in J6.1. This justification explains how the scenarios were identified such that they represent real scenarios of interest that may be encountered when the system is in operation.

The strategy to support the integration claim is to firstly use the integration test results ([FF]) to demonstrate the safety requirements are met for the defined operating scenarios. Integration testing is often performed for autonomous systems using a simulator. Where this is the case it is also necessary to demonstrate that the simulations that are used are a sufficient representation of the operational system to which the ML component is deployed. Evidence for this will be provided to support claim G6.5.

It must also be demonstrated that the safety requirements allocated to the ML component continue to be satisfied during the operation of the system. To demonstrate this, claim G6.6 shows that the system is designed such that it supports the safe operation of the ML component, and G6.7 demonstrates that the observed behaviour during operation continues to satisfy the safety requirements. In a complete safety case for an ML component argument and evidence to support this claim would be required, further guidance on this is provided in [6].

It must be demonstrated that the design of the system into which the ML component is integrated is robust by taking account of the identified potential erroneous behaviour ([DD]). It must be shown that predicted erroneous behaviour will not result in violation of the safety requirements. In particular the argument must focus on erroneous inputs to the ML component from the rest of the system and erroneous outputs from the ML component itself. The argument must also consider assumptions made about the system and the operating environment during the development of the ML component that may become invalid during operation. The sufficiency of the identification of these erroneous behaviours must be justified in J6.2. This may be informed by the results of system safety analysis activities. Claim G6.6 is supported by two sub‐claims, one that demonstrates the system design incorporates sufficient monitoring of erroneous behaviours, and one demonstrating that the response of the system to such behaviours is acceptable.

It must be demonstrated that the system design incorporates sufficient monitoring of the identified erroneous behaviour to ensure that any behaviour that could result in violation of a safety requirement will be identified if it occurs during operation.

It must be demonstrated that the system design ensures that an acceptable response can be provided if monitoring reveals erroneous behaviour during operation. The response may take many forms, depending on the nature of the system, the relevant system hazard behaviour and the erroneous behaviour identified. This may include, for example, the provision of redundancy in the system architecture or the specification of safe degraded operation. Evidence should be provided to show that a sufficiently safe response is provided.

Continue to: Stage 7. Afterword - authors and thanks

Our site depends on cookies to provide our service to you. If you continue to use this site we will assume that you are happy with that. View our privacy policy.