The Winner With Nothing Left to Win

Retrieval check (from lesson 3, new setting). A thermostat runs a tiny production system — condition→action rules over a working memory of facts, picking one matched rule per cycle (recall lesson 3: this is the recognize–act cycle). Working memory: {room: cold, heater: off, fan: off}. Rules: R1: IF room=cold AND heater=off THEN heater=on. R2: IF heater=on AND fan=off THEN fan=on. R3: IF room=cold THEN log-reading (matches but changes nothing visible; conflict resolution favors the most specific — most-conditions — match). In what order do rules fire, and how many cycles run before nothing new matches? Work it out before reading on — the solution confirms it.

Back to the blackboard

Lesson 6 built a blackboard — a shared workspace holding hypotheses at multiple levels, each with a confidence — worked on by independent knowledge sources (KSs), specialist modules that fire when their trigger condition matches something currently on the board. Whenever a KS could fire, a KSAR (knowledge-source activation record) is queued on the agenda: “KS X could fire on hypothesis Y.” Control’s job, every cycle, is to score the pending KSARs and run the best one — focus of attention.

Lesson 6 used the heuristic priority = credibility ÷ cost: favor actions that are both likely to be right and cheap to try. That heuristic has a real failure mode, and it’s not the one it looks like.

The pathology. A speech-understanding system has an island hypothesis — a stretch it’s already 99% confident is the word “seven” — sitting at confidence 0.99. Three KSARs are pending:

Score credibility ÷ cost: K1 = 0.95, K2 = 0.30, K3 = 0.33. K1 wins, every cycle, forever — it’s cheap and the KS that runs it is reliable. So the scheduler spends cycle after cycle re-verifying a hypothesis that was already essentially certain, while two seconds of the utterance never gets looked at. (This is the same pattern as a linter that keeps re-running on a file that already passes every rule, because “cheap and reliable” scores well regardless of whether there’s anything left to find.)

The tempting wrong fix: “the problem is repetition — raise K1’s cost every time it runs, so it stops winning.” That patches this KSAR, this cycle. It does nothing about K3 doing the same thing next week on a different island, or about a KS whose action is genuinely free to run (cost ≈ 0) but still contributes nothing. Cost was never the broken variable.

What credibility ÷ cost is missing: a value term — the expected change this action would make to the global interpretation. Re-verifying a 0.99 island can, at best, nudge it to maybe 0.995 — tiny expected value, regardless of cost. Covering two seconds of currently-unexplained signal, by contrast, has high value: it’s the difference between “no hypothesis” and “some hypothesis,” the largest kind of improvement the board can register.

Revised scoring: priority = credibility × value ÷ cost, where value is the expected gain in solution quality if the action succeeds:

Compute priority = credibility × value ÷ cost for all three, and report the winning score.