Thermostat

A continuous-state Markov decision process for HVAC control. A room's temperature is a floating-point measure, not a discrete state. Each cycle the controller heats if the temperature is at or below the threshold and coasts otherwise; the heater adds a normally-distributed gain and the weather subtracts a normally-distributed loss, so the trajectory drifts continuously. Comfort and energy are tracked as separate reward channels.

Shows the engine handling continuous state and distribution-valued updates, not just discrete transitions. The two independent reward channels let you study the comfort-versus-energy trade-off offline, directly from the exported room rows.

# thermostat.osi.yaml — emitted automatically
semantic_model:
  name: "thermostat"
  source: "duckdb://thermostat.db"
  entities:
    - name: room
      primary_key: id
  dimensions:
    - name: state
      type: categorical
    - name: t
      type: time
  measures:
    - name: row_count
      agg: count
    - name: active
      agg: sum
      filter: "state = 'ACTIVE'"