Subjective machines: Probabilistic risk assessment based on deep learning of soft information

Area of Science:

• Artificial Intelligence
• Risk Management
• Autonomous Systems

Background:

• Machine learning (ML) is established for risk classification, failure diagnosis, and condition monitoring.
• ML has not been widely applied to probabilistic risk assessment (PRA).
• Expert judgments in PRA are subjective and prone to biases.

Purpose of the Study:

• To develop natural language-based PRA models using deep learning.
• To emulate expert quantified risk estimates.
• To enable a priori risk assessment with limited text and numeric data.

Main Methods:

• Applied deep learning algorithms to natural language data.
• Utilized Universal Sentence Embedding (USE) and Gradient Boosting Regression (GBR) trees.
• Trained models on limited structured data.

Main Results:

• The USE with GBR approach showed promising results.
• Generated survival distributions for autonomous system loss likelihood.
• No statistically significant difference found between ML and expert results in open water/ice shelf environments.

Conclusions:

• Natural language-based ML models can effectively emulate expert risk assessments.
• This approach provides a viable method for a priori probabilistic risk assessment.
• The developed models show high accuracy comparable to human experts.