Precrime - Self-assessment Oracles for Anticipatory Testing

One of the long-lasting dreams in science fiction is the ability to arrest criminals before they even commit crimes. Software testing researchers have a similar dream: when the context for a bug manifestation occurs in the field, the goal is to discover and fix the bug before it causes any in-field failure. In fact, current practice of pre-release testing is severely limited when dealing with autonomous AI (Artificial Intelligence) systems (such as self-driving cars, robots, automated traders, virtual doctors and customer service chatbots), running in complex, rapidly changing environments, which cause their run-time adaptation, learning and knowledge acquisition, because pre-release testing cannot exhaustively explore all different contexts and states in which the software will be running.

The PRECRIME project introduces a new, disruptive view on testing, called anticipatory testing and aimed at fixing bugs before they even manifest themselves in the field. Anticipatory testing is activated at run-time by a new type of oracles, called self-assessment oracles, which observe and report unexpected execution contexts. A self-assessment oracle is an estimator of the system’s confidence in being able to handle a new execution context correctly. The goal of anticipatory testing is to anticipate any failure that might occur in the field due to unexpected execution contexts. Whenever an execution context monitored at runtime by self-assessment oracles is estimated as a low confidence context for the system, anticipatory testing exercises the software automatically and extensively in similar contexts. Timely activation of anticipatory testing by runtime observations results in early, anticipatory fault detection. Combined with automated patch synthesis, anticipatory testing leads to the release of a patch for the fault before any software failure occurs in the field.