fix(ai): prevent greedy engine from freezing after a high-value ability activation
When an ability activation raised the score baseline significantly (e.g. bouncing a unit, dealing damage, healing), no subsequent move or attack could beat that new baseline, so the AI ended its turn after just one action leaving all other units idle. Two-part fix: 1. Relax the selection threshold by NoRegressTolerance (1.0 pts) so that moves/attacks that are near-neutral still qualify as candidates. 2. Track serialised-state hashes for every state seen during the turn; candidates that would revisit a prior state are skipped, preventing the tolerance from enabling intra-turn cycles. Generated with [Claude Code](https://claude.ai/code) via [Happy](https://happy.engineering) Co-Authored-By: Claude <noreply@anthropic.com> Co-Authored-By: Happy <yesreply@happy.engineering>
This commit is contained in:
@@ -52,6 +52,18 @@ internal sealed class HeuristicDecisionEngine : IAiDecisionEngine
|
||||
const int maxActions = 50;
|
||||
int iterations = 0;
|
||||
|
||||
// Track serialised-state hashes seen so far this turn to prevent intra-turn
|
||||
// cycles. Using full JSON hash is cheap enough given we already serialise for
|
||||
// every candidate evaluation.
|
||||
var seenStateHashes = new HashSet<int> { HashState(workingState) };
|
||||
|
||||
// Allow actions that don't strictly beat the current baseline but are no
|
||||
// worse than this tolerance. This prevents the AI from freezing after a
|
||||
// single high-value action (e.g. an ability activation) raises the baseline
|
||||
// above what any subsequent move/attack can achieve, leaving other units idle.
|
||||
// Cycle detection (seenStateHashes) ensures the tolerance cannot cause loops.
|
||||
const float NoRegressTolerance = 1.0f;
|
||||
|
||||
while (iterations++ < maxActions)
|
||||
{
|
||||
var workingCtx = new AiTurnContext(workingState, ctx.AiAccountId);
|
||||
@@ -61,17 +73,20 @@ internal sealed class HeuristicDecisionEngine : IAiDecisionEngine
|
||||
|
||||
if (candidates.Count == 0) break;
|
||||
|
||||
AiAction? bestAction = null;
|
||||
float bestScore = baseScore;
|
||||
AiAction? bestAction = null;
|
||||
float bestScore = baseScore - NoRegressTolerance; // threshold: must not regress beyond tolerance
|
||||
int bestNewHash = 0;
|
||||
|
||||
foreach (var candidate in candidates)
|
||||
{
|
||||
var simState = CloneState(workingState);
|
||||
float score;
|
||||
int newHash;
|
||||
try
|
||||
{
|
||||
ApplyAction(simState, ctx.AiAccountId, candidate);
|
||||
score = ScoringFunctions.Evaluate(simState, ctx.AiAccountId, _profile);
|
||||
newHash = HashState(simState);
|
||||
score = ScoringFunctions.Evaluate(simState, ctx.AiAccountId, _profile);
|
||||
}
|
||||
catch
|
||||
{
|
||||
@@ -79,19 +94,26 @@ internal sealed class HeuristicDecisionEngine : IAiDecisionEngine
|
||||
continue;
|
||||
}
|
||||
|
||||
// Skip actions that would revisit a state already seen this turn.
|
||||
if (seenStateHashes.Contains(newHash)) continue;
|
||||
|
||||
if (score > bestScore)
|
||||
{
|
||||
bestScore = score;
|
||||
bestAction = candidate;
|
||||
bestScore = score;
|
||||
bestAction = candidate;
|
||||
bestNewHash = newHash;
|
||||
}
|
||||
}
|
||||
|
||||
if (bestAction is null)
|
||||
break; // no improvement available
|
||||
break; // no improvement or non-regressive action available
|
||||
|
||||
// ── Mistake injection ──────────────────────────────────────────────
|
||||
if (_profile.MistakeRate > 0 && Random.Shared.NextSingle() < _profile.MistakeRate)
|
||||
bestAction = candidates[Random.Shared.Next(candidates.Count)];
|
||||
{
|
||||
bestAction = candidates[Random.Shared.Next(candidates.Count)];
|
||||
bestNewHash = 0; // hash unknown for random pick; recomputed below
|
||||
}
|
||||
|
||||
plan.Add(bestAction);
|
||||
|
||||
@@ -99,6 +121,8 @@ internal sealed class HeuristicDecisionEngine : IAiDecisionEngine
|
||||
try { ApplyAction(workingState, ctx.AiAccountId, bestAction); }
|
||||
catch { break; /* action became invalid — stop the loop */ }
|
||||
|
||||
seenStateHashes.Add(bestNewHash != 0 ? bestNewHash : HashState(workingState));
|
||||
|
||||
baseScore = ScoringFunctions.Evaluate(workingState, ctx.AiAccountId, _profile);
|
||||
}
|
||||
|
||||
@@ -111,6 +135,9 @@ internal sealed class HeuristicDecisionEngine : IAiDecisionEngine
|
||||
private static GameState CloneState(GameState state)
|
||||
=> GameEngine.Deserialize(GameEngine.Serialize(state));
|
||||
|
||||
private static int HashState(GameState state)
|
||||
=> GameEngine.Serialize(state).GetHashCode();
|
||||
|
||||
/// <summary>
|
||||
/// Apply one AiAction directly to a GameState (for simulation purposes).
|
||||
/// Throws GameException if the action is invalid — callers must handle this.
|
||||
|
||||
Reference in New Issue
Block a user