An active inference approach to on-line agent monitoring in safety-critical systems

AVILA, LUIS OMAR; MARTINEZ, ERNESTO CARLOS

ADVANCED ENGINEERING INFORMATICS

ELSEVIER SCI LTD

Lugar: Amsterdam; Año: 2015 vol. 29 p. 1083 - 1095

1474-0346

The current trend towards integrating software agents in safety-critical systems such asdrones, autonomous cars and medical devices, which must operate in uncertainenvironments, gives rise to the need of on-line detection of an unexpected behavior. In thiswork, on-line monitoring is carried out by comparing environmental state transitions withprior beliefs descriptive of optimal behavior. The agent policy is computed analyticallyusing linearly solvable Markov decision processes. Active inference using prior beliefsallows a monitor proactively rehearsing on-line future agent actions over a rolling horizonso as to generate expectations to discover surprising behaviors. A Bayesian surprise metricis proposed based on twin Gaussian processes to measure the difference between prior andposterior beliefs about state transitions in the agent environment. Using a sliding window ofsampled data, beliefs are updated a posteriori by comparing a sequence of state transitionswith the ones predicted using the optimal policy. An artificial pancreas for diabetic patientsis used as a representative example.