Situational Control

Origin. Dmitry Pospelov and the Soviet AI school, Institute of Control Sciences (IPU), Moscow, 1970s-1980s. Developed as a distinct approach from Western AI's focus on search and logic. Emphasized knowledge-intensive reasoning about structured wholes rather than state-space search.

Mechanism. The world is represented as structured situations rather than states in a state space. A situation is a configuration of objects, relations, and context that has meaning as a whole; the same objects in different relations constitute different situations. Control proceeds by recognizing the current situation type, retrieving the appropriate response schema, and instantiating it. Expertise is encoded as situation-action pairs, not as goal-means chains. The method is knowledge-intensive rather than search-intensive.

Procedure. Design and implement situational control systems: (1) Enumerate situation types — what distinct situations must the system handle? Each type is a recognizable configuration with its own appropriate response. Work with domain experts to elicit the situation vocabulary. (2) Define recognition criteria — for each situation type, what observable features identify it? What configuration of sensor data, context, and history signals this situation? Recognition criteria should be robust to noise. (3) Define response schemas — for each situation type, what actions are appropriate? In what sequence? With what parameters? What adaptations based on situation specifics? (4) Implement matching — when new observations arrive, match against situation types. Use pattern matching, classification, or rule-based recognition. (5) Handle boundaries — situations near type boundaries may match multiple types or none. Design default behaviors for ambiguous cases and escalation paths for unrecognized situations. (6) Execute and adapt — on match, instantiate and execute the response schema. Monitor outcomes to refine recognition criteria and responses.

Applies to. Expert systems. Diagnostic systems. Robotics. Operational procedures. Any domain where expertise consists of recognizing situations and applying appropriate responses — troubleshooting, medical diagnosis, process control.

Limitations. Brittle at situation boundaries: real situations are continuous, but the method discretizes them into types. Observations near boundaries produce unstable recognition. The enumeration of situation types is complete only for well-understood domains; novel situations fall through. The method encodes expertise but does not generate it — it cannot handle situations the designer did not anticipate. Knowledge acquisition from experts is labor-intensive.