Troubleshooting

This guide covers the failure modes and operational pathologies that are most likely to appear when integrating timetable-sa into a real optimization workflow. Each section explains the observed symptom, the code-level cause, and the most reliable corrective actions.

solve() throws SolveConcurrencyError

This error occurs when a second solve() call starts before the previous call on the same solver instance has finished.

Why it happens

SimulatedAnnealing<TState> uses an instance-level isSolving guard. The guard is checked at the start of solve() and cleared in a finally block.

How to fix it

• await the current solve before calling solve() again,
• create one solver instance per concurrent run,
• avoid sharing a single instance across overlapping async workflows.

Safe pattern

const solverA = new SimulatedAnnealing(state, constraints, moves, config);
const solverB = new SimulatedAnnealing(state, constraints, moves, config);

const [solutionA, solutionB] = await Promise.all([
  solverA.solve(),
  solverB.solve(),
]);

SAConfigError during construction

SAConfigError indicates that the constructor rejected the supplied state, constraint list, move generators, or configuration.

Common causes

• initialState is null or undefined,
• a constraint is missing name, type, or evaluate,
• a move generator is missing name, generate, or canApply,
• coolingRate is not strictly between 0 and 1,
• maxIterations is not a positive integer,
• an optional numeric field violates its integer or range constraint,
• a soft constraint has a negative weight.

Important non-causes

The current validator does not reject these conditions by itself:

• empty constraints arrays,
• empty moveGenerators arrays,
• minTemperature >= initialTemperature.

If your application requires these invariants, enforce them at the integration layer.

ConstraintValidationError at runtime

ConstraintValidationError indicates that a constraint returned an invalid score during evaluation.

Exact trigger

The engine throws this error when evaluate(state) returns:

• a non-finite value such as NaN or Infinity, or
• a finite value outside the interval [0, 1].

Common misunderstanding

The score is a satisfaction score, not a violation score.

• 1 means satisfied,
• 0 means violated,
• intermediate values mean partial satisfaction.

Returning higher numbers for worse states is a common integration mistake.

Recommended fix

Normalize every constraint explicitly.

evaluate(state) {
  const raw = computeSatisfaction(state);
  return Math.max(0, Math.min(1, raw));
}

Constraint evaluation throws a custom error

If your evaluate() function throws its own exception, that exception normally propagates directly. The engine does not automatically wrap arbitrary user code errors into ConstraintValidationError.

How to handle it

• catch and rethrow with domain-specific context inside the constraint, or
• validate required state invariants before running the solver.

No progress callbacks are arriving

If onProgress is configured but you do not see updates, the issue is often cadence-related rather than callback-related.

How callbacks are scheduled

The solver emits progress:

• at iteration 0,
• every logging.logInterval iterations,
• on forced events such as reheating.

Checklist

• confirm that onProgress is actually defined,
• reduce logging.logInterval if callbacks are too sparse,
• confirm that the solve is reaching enough iterations,
• confirm that all move generators are not immediately inapplicable.

Progress callback is slow

If throughput drops dramatically after enabling onProgress, the callback is likely dominating wall-clock time.

Why it happens

By default, onProgressMode is 'await', which means the solver waits for the callback to finish before continuing.

How to fix it

• switch to onProgressMode: 'fire-and-forget',
• reduce callback side effects,
• batch external writes,
• increase logging.logInterval to reduce callback frequency.

Important implementation detail

The callback receives state = null, not a cloned state snapshot. If your code expects a real state object, it may silently fail or misbehave.

Progress callback throws errors, but solve continues

This behavior is expected.

Why it happens

ProgressReporter catches callback errors and forwards them to the logger as warnings. The optimization loop continues because progress telemetry is treated as observational, not correctness-critical.

How to debug it

• enable logging.level: 'warn' or 'debug',
• inspect the callback for rejected promises,
• add explicit application-side error capture if telemetry failure must abort the run.

File logging is not created

If log files are missing, verify the configuration and the runtime environment.

Checklist

• set logging.output to 'file' or 'both',
• set a writable logging.filePath,
• confirm the process has permission to write to the target location,
• confirm logging is not disabled with logging.enabled: false.

Implementation detail

The logger creates missing parent directories automatically. If the file still does not appear, the problem is usually permissions or an unexpected working directory.

Solver exits early with very few iterations

Early termination often indicates that the search cannot generate valid neighbors or that one of the stopping conditions is reached much sooner than expected.

Likely causes

• all move generators return false from canApply(...),
• minTemperature is too high for the chosen coolingRate,
• maxIterations is too low,
• Phase 1 reaches zero hard violations quickly and Phase 2 cools out fast.

How to investigate

• add logging at info or debug level,
• inspect whether any operator attempts are recorded in operatorStats,
• verify that move generators remain applicable across the full state space.

Phase 1.5 ends too quickly

If intensification appears to trigger and then stop almost immediately, the most common causes are the explicit Phase 1.5 budget cap or the per-attempt early-stop threshold.

How to investigate

Inspect either solution.diagnostics.intensification or solver.getDiagnostics().intensification after the run.

• phase15BudgetLimitIterations shows the global Phase 1.5 budget.
• phase15BudgetUsedIterations shows how much of that budget was consumed.
• phase15EndedByBudget tells you the phase exhausted the budget.
• phase15EndedByEarlyStop tells you the current attempt ended because the global best hard-violation objective stalled.

How to fix it

• increase intensificationBudgetFractionOfMaxIterations when the phase ends by budget too early,
• increase intensificationEarlyStopNoBestImproveIterations when the search needs more patience,
• increase maxIterations if the Phase 1.5 budget is too small because the global run budget itself is too small.

Intensification does not use the operators you expect

If Phase 1.5 keeps choosing broad exploratory moves instead of the repair operators you intended, verify the explicit targeting configuration first.

Why it happens

The branch implementation uses intensificationTargetedOperatorNames as a case-insensitive exact-name match. If the names do not match, the targeted set is empty and Phase 1.5 falls back to all applicable generators.

Tabu gating can also make a good operator appear inactive if its candidates are being skipped before acceptance.

How to fix it

• compare your configured names against moveGenerator.name exactly,
• increase intensificationTargetedSelectionRate if the targeted set exists but is chosen too rarely,
• inspect phase15TabuSkips in diagnostics,
• disable intensificationUseTabu temporarily when Phase 1.5 becomes too conservative during hard-repair search.

Poor convergence or high variance

High variance across runs is expected to some degree because the solver uses Math.random() internally. Excessive variance usually points to weak move design, weak scoring gradients, or unstable configuration.

Corrective actions

• increase maxIterations,
• cool more slowly by moving coolingRate closer to 1,
• increase hardConstraintWeight if hard feasibility is inconsistent,
• enable tabu search and intensification,
• add more targeted move generators,
• ensure constraints return informative partial-satisfaction scores instead of only binary outputs where graded feedback is possible.

Hard violations do not drop effectively

When hard violations remain stubbornly high, the problem is often not the anneal schedule alone.

Likely causes

• hardConstraintWeight is too small,
• move generators do not directly repair the dominant violations,
• operator names do not align with the Phase 1 targeting heuristics,
• getViolations() is missing, so hard-violation multiplicity is only inferred,
• the problem instance may be infeasible.

Recommended actions

• increase hardConstraintWeight,
• add or rename repair-oriented operators with names like fix, swap, change, capacity, lecturer, or domain-specific equivalents,
• implement getViolations() for hard constraints,
• run several independent solves before concluding infeasibility.

Tabu search does not help

When enabling tabu produces little improvement, the state-signature function is usually the first thing to inspect.

Likely causes

• signatures collide because they ignore meaningful state differences,
• signatures are too expensive to compute and add overhead without enough value,
• the tenure is too short to suppress cycling,
• the underlying move set already has low cycling risk.

Fixes

• implement a domain-specific getStateSignature(...),
• increase tabuTenure moderately,
• inspect tabuHits through ProgressStats to confirm the feature is active.

Tabu search fails with signature-generation errors

If the solver cannot create a deterministic default signature and no custom signature function is provided, it throws a plain Error asking for config.getStateSignature.

When this happens

This is most common when the state contains unusual object graphs or structures that cannot be serialized deterministically by the fallback path.

Fix

Provide a stable signature that captures only the optimization-relevant parts of the state.

Solution quality looks inconsistent with softViolations

This is usually a reporting misunderstanding.

Important distinction

• fitness is the weighted optimization objective,
• softViolations in Solution<TState> is a count of soft-violation records,
• ProgressStats.softViolations is also a count, not a weighted soft penalty.

Two solutions can therefore have the same softViolations count but very different fitness values.

Final recommendation

When debugging difficult behavior, inspect the system in this order:

1. verify score semantics and clone correctness,
2. verify move-generator applicability and diversity,
3. verify state-signature quality for tabu,
4. tune annealing and intensification parameters,
5. compare several runs before drawing conclusions from a single trajectory.

Next steps

If you need more implementation detail after debugging:

• read API Reference for precise type and error contracts,
• read Configuration for tuning strategy,
• read Algorithm and Runtime Behavior for the algorithm rules behind these symptoms.