@arcanea/guardian-evolution 0.1.0

package/dist/algorithms/ppo.js

@@ -0,0 +1,331 @@
+/**
+ * Proximal Policy Optimization (PPO)
+ *
+ * Implements PPO algorithm for stable policy learning with:
+ * - Clipped surrogate objective
+ * - GAE (Generalized Advantage Estimation)
+ * - Value function clipping
+ * - Entropy bonus
+ *
+ * Performance Target: <10ms per update step
+ */
+/**
+ * Default PPO configuration
+ */
+export const DEFAULT_PPO_CONFIG = {
+    algorithm: 'ppo',
+    learningRate: 0.0003,
+    gamma: 0.99,
+    entropyCoef: 0.01,
+    valueLossCoef: 0.5,
+    maxGradNorm: 0.5,
+    epochs: 4,
+    miniBatchSize: 64,
+    clipRange: 0.2,
+    clipRangeVf: null,
+    targetKL: 0.01,
+    gaeLambda: 0.95,
+};
+/**
+ * PPO Algorithm Implementation
+ */
+export class PPOAlgorithm {
+    config;
+    // Policy network weights (simplified linear model for speed)
+    policyWeights;
+    valueWeights;
+    // Optimizer state
+    policyMomentum;
+    valueMomentum;
+    // Experience buffer
+    buffer = [];
+    // Statistics
+    updateCount = 0;
+    totalLoss = 0;
+    approxKL = 0;
+    clipFraction = 0;
+    constructor(config = {}) {
+        this.config = { ...DEFAULT_PPO_CONFIG, ...config };
+        // Initialize weights (768 input dim, simplified)
+        const dim = 768;
+        this.policyWeights = new Float32Array(dim);
+        this.valueWeights = new Float32Array(dim);
+        this.policyMomentum = new Float32Array(dim);
+        this.valueMomentum = new Float32Array(dim);
+        // Xavier initialization
+        const scale = Math.sqrt(2 / dim);
+        for (let i = 0; i < dim; i++) {
+            this.policyWeights[i] = (Math.random() - 0.5) * scale;
+            this.valueWeights[i] = (Math.random() - 0.5) * scale;
+        }
+    }
+    /**
+     * Add experience from trajectory
+     */
+    addExperience(trajectory) {
+        if (trajectory.steps.length === 0)
+            return;
+        // Compute values for each step
+        const values = trajectory.steps.map(step => this.computeValue(step.stateAfter));
+        // Compute advantages using GAE
+        const advantages = this.computeGAE(trajectory.steps.map(s => s.reward), values);
+        // Compute returns
+        const returns = this.computeReturns(trajectory.steps.map(s => s.reward));
+        // Add to buffer
+        for (let i = 0; i < trajectory.steps.length; i++) {
+            const step = trajectory.steps[i];
+            this.buffer.push({
+                state: step.stateAfter,
+                action: this.hashAction(step.action),
+                reward: step.reward,
+                value: values[i],
+                logProb: this.computeLogProb(step.stateAfter, step.action),
+                advantage: advantages[i],
+                return_: returns[i],
+            });
+        }
+    }
+    /**
+     * Perform PPO update
+     * Target: <10ms
+     */
+    update() {
+        const startTime = performance.now();
+        if (this.buffer.length < this.config.miniBatchSize) {
+            return { policyLoss: 0, valueLoss: 0, entropy: 0 };
+        }
+        // Normalize advantages
+        const advantages = this.buffer.map(e => e.advantage);
+        const advMean = advantages.reduce((a, b) => a + b, 0) / advantages.length;
+        const advStd = Math.sqrt(advantages.reduce((a, b) => a + (b - advMean) ** 2, 0) / advantages.length) + 1e-8;
+        for (const exp of this.buffer) {
+            exp.advantage = (exp.advantage - advMean) / advStd;
+        }
+        let totalPolicyLoss = 0;
+        let totalValueLoss = 0;
+        let totalEntropy = 0;
+        let totalClipFrac = 0;
+        let totalKL = 0;
+        let numUpdates = 0;
+        // Multiple epochs
+        for (let epoch = 0; epoch < this.config.epochs; epoch++) {
+            // Shuffle buffer
+            this.shuffleBuffer();
+            // Process mini-batches
+            for (let i = 0; i < this.buffer.length; i += this.config.miniBatchSize) {
+                const batch = this.buffer.slice(i, i + this.config.miniBatchSize);
+                if (batch.length < this.config.miniBatchSize / 2)
+                    continue;
+                const result = this.updateMiniBatch(batch);
+                totalPolicyLoss += result.policyLoss;
+                totalValueLoss += result.valueLoss;
+                totalEntropy += result.entropy;
+                totalClipFrac += result.clipFrac;
+                totalKL += result.kl;
+                numUpdates++;
+                // Early stopping if KL too high
+                if (result.kl > this.config.targetKL * 1.5) {
+                    break;
+                }
+            }
+        }
+        // Clear buffer
+        this.buffer = [];
+        this.updateCount++;
+        const elapsed = performance.now() - startTime;
+        if (elapsed > 10) {
+            console.warn(`PPO update exceeded target: ${elapsed.toFixed(2)}ms > 10ms`);
+        }
+        return {
+            policyLoss: numUpdates > 0 ? totalPolicyLoss / numUpdates : 0,
+            valueLoss: numUpdates > 0 ? totalValueLoss / numUpdates : 0,
+            entropy: numUpdates > 0 ? totalEntropy / numUpdates : 0,
+        };
+    }
+    /**
+     * Get action from policy
+     */
+    getAction(state) {
+        const logits = this.computeLogits(state);
+        const probs = this.softmax(logits);
+        const action = this.sampleAction(probs);
+        return {
+            action,
+            logProb: Math.log(probs[action] + 1e-8),
+            value: this.computeValue(state),
+        };
+    }
+    /**
+     * Get statistics
+     */
+    getStats() {
+        return {
+            updateCount: this.updateCount,
+            bufferSize: this.buffer.length,
+            avgLoss: this.updateCount > 0 ? this.totalLoss / this.updateCount : 0,
+            approxKL: this.approxKL,
+            clipFraction: this.clipFraction,
+        };
+    }
+    // ==========================================================================
+    // Private Methods
+    // ==========================================================================
+    computeValue(state) {
+        let value = 0;
+        for (let i = 0; i < Math.min(state.length, this.valueWeights.length); i++) {
+            value += state[i] * this.valueWeights[i];
+        }
+        return value;
+    }
+    computeLogits(state) {
+        // Simplified: 4 discrete actions
+        const numActions = 4;
+        const logits = new Float32Array(numActions);
+        for (let a = 0; a < numActions; a++) {
+            for (let i = 0; i < Math.min(state.length, this.policyWeights.length); i++) {
+                logits[a] += state[i] * this.policyWeights[i] * (1 + a * 0.1);
+            }
+        }
+        return logits;
+    }
+    computeLogProb(state, action) {
+        const logits = this.computeLogits(state);
+        const probs = this.softmax(logits);
+        const actionIdx = this.hashAction(action);
+        return Math.log(probs[actionIdx] + 1e-8);
+    }
+    hashAction(action) {
+        // Simple hash to action index (0-3)
+        let hash = 0;
+        for (let i = 0; i < action.length; i++) {
+            hash = (hash * 31 + action.charCodeAt(i)) % 4;
+        }
+        return hash;
+    }
+    softmax(logits) {
+        const max = Math.max(...logits);
+        const exps = new Float32Array(logits.length);
+        let sum = 0;
+        for (let i = 0; i < logits.length; i++) {
+            exps[i] = Math.exp(logits[i] - max);
+            sum += exps[i];
+        }
+        for (let i = 0; i < exps.length; i++) {
+            exps[i] /= sum;
+        }
+        return exps;
+    }
+    sampleAction(probs) {
+        const r = Math.random();
+        let cumSum = 0;
+        for (let i = 0; i < probs.length; i++) {
+            cumSum += probs[i];
+            if (r < cumSum)
+                return i;
+        }
+        return probs.length - 1;
+    }
+    computeGAE(rewards, values) {
+        const advantages = new Array(rewards.length).fill(0);
+        let lastGae = 0;
+        for (let t = rewards.length - 1; t >= 0; t--) {
+            const nextValue = t < rewards.length - 1 ? values[t + 1] : 0;
+            const delta = rewards[t] + this.config.gamma * nextValue - values[t];
+            lastGae = delta + this.config.gamma * this.config.gaeLambda * lastGae;
+            advantages[t] = lastGae;
+        }
+        return advantages;
+    }
+    computeReturns(rewards) {
+        const returns = new Array(rewards.length).fill(0);
+        let cumReturn = 0;
+        for (let t = rewards.length - 1; t >= 0; t--) {
+            cumReturn = rewards[t] + this.config.gamma * cumReturn;
+            returns[t] = cumReturn;
+        }
+        return returns;
+    }
+    shuffleBuffer() {
+        for (let i = this.buffer.length - 1; i > 0; i--) {
+            const j = Math.floor(Math.random() * (i + 1));
+            [this.buffer[i], this.buffer[j]] = [this.buffer[j], this.buffer[i]];
+        }
+    }
+    updateMiniBatch(batch) {
+        let policyLoss = 0;
+        let valueLoss = 0;
+        let entropy = 0;
+        let clipFrac = 0;
+        let kl = 0;
+        const policyGrad = new Float32Array(this.policyWeights.length);
+        const valueGrad = new Float32Array(this.valueWeights.length);
+        for (const exp of batch) {
+            // Current policy
+            const logits = this.computeLogits(exp.state);
+            const probs = this.softmax(logits);
+            const newLogProb = Math.log(probs[exp.action] + 1e-8);
+            const currentValue = this.computeValue(exp.state);
+            // Ratio for PPO
+            const ratio = Math.exp(newLogProb - exp.logProb);
+            // Clipped surrogate objective
+            const surr1 = ratio * exp.advantage;
+            const surr2 = Math.max(Math.min(ratio, 1 + this.config.clipRange), 1 - this.config.clipRange) * exp.advantage;
+            const policyLossI = -Math.min(surr1, surr2);
+            policyLoss += policyLossI;
+            // Track clipping
+            if (Math.abs(ratio - 1) > this.config.clipRange) {
+                clipFrac++;
+            }
+            // KL divergence approximation
+            kl += (exp.logProb - newLogProb);
+            // Value loss
+            let valueLossI;
+            if (this.config.clipRangeVf !== null) {
+                const valuePred = currentValue;
+                const valueClipped = exp.value + Math.max(Math.min(valuePred - exp.value, this.config.clipRangeVf), -this.config.clipRangeVf);
+                const vf1 = (valuePred - exp.return_) ** 2;
+                const vf2 = (valueClipped - exp.return_) ** 2;
+                valueLossI = Math.max(vf1, vf2);
+            }
+            else {
+                valueLossI = (currentValue - exp.return_) ** 2;
+            }
+            valueLoss += valueLossI;
+            // Entropy
+            let entropyI = 0;
+            for (const p of probs) {
+                if (p > 0)
+                    entropyI -= p * Math.log(p);
+            }
+            entropy += entropyI;
+            // Compute gradients (simplified)
+            for (let i = 0; i < Math.min(exp.state.length, policyGrad.length); i++) {
+                policyGrad[i] += exp.state[i] * policyLossI * 0.01;
+                valueGrad[i] += exp.state[i] * valueLossI * 0.01;
+            }
+        }
+        // Apply gradients with momentum
+        const lr = this.config.learningRate;
+        const beta = 0.9;
+        for (let i = 0; i < this.policyWeights.length; i++) {
+            this.policyMomentum[i] = beta * this.policyMomentum[i] + (1 - beta) * policyGrad[i];
+            this.policyWeights[i] -= lr * this.policyMomentum[i];
+            this.valueMomentum[i] = beta * this.valueMomentum[i] + (1 - beta) * valueGrad[i];
+            this.valueWeights[i] -= lr * this.valueMomentum[i];
+        }
+        return {
+            policyLoss: policyLoss / batch.length,
+            valueLoss: valueLoss / batch.length,
+            entropy: entropy / batch.length,
+            clipFrac: clipFrac / batch.length,
+            kl: kl / batch.length,
+        };
+    }
+}
+/**
+ * Factory function
+ */
+export function createPPO(config) {
+    return new PPOAlgorithm(config);
+}
+//# sourceMappingURL=ppo.js.map

package/dist/algorithms/ppo.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ppo.js","sourceRoot":"","sources":["../../src/algorithms/ppo.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAQH;;GAEG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAc;IAC3C,SAAS,EAAE,KAAK;IAChB,YAAY,EAAE,MAAM;IACpB,KAAK,EAAE,IAAI;IACX,WAAW,EAAE,IAAI;IACjB,aAAa,EAAE,GAAG;IAClB,WAAW,EAAE,GAAG;IAChB,MAAM,EAAE,CAAC;IACT,aAAa,EAAE,EAAE;IACjB,SAAS,EAAE,GAAG;IACd,WAAW,EAAE,IAAI;IACjB,QAAQ,EAAE,IAAI;IACd,SAAS,EAAE,IAAI;CAChB,CAAC;AAeF;;GAEG;AACH,MAAM,OAAO,YAAY;IACf,MAAM,CAAY;IAE1B,6DAA6D;IACrD,aAAa,CAAe;IAC5B,YAAY,CAAe;IAEnC,kBAAkB;IACV,cAAc,CAAe;IAC7B,aAAa,CAAe;IAEpC,oBAAoB;IACZ,MAAM,GAAoB,EAAE,CAAC;IAErC,aAAa;IACL,WAAW,GAAG,CAAC,CAAC;IAChB,SAAS,GAAG,CAAC,CAAC;IACd,QAAQ,GAAG,CAAC,CAAC;IACb,YAAY,GAAG,CAAC,CAAC;IAEzB,YAAY,SAA6B,EAAE;QACzC,IAAI,CAAC,MAAM,GAAG,EAAE,GAAG,kBAAkB,EAAE,GAAG,MAAM,EAAE,CAAC;QAEnD,iDAAiD;QACjD,MAAM,GAAG,GAAG,GAAG,CAAC;QAChB,IAAI,CAAC,aAAa,GAAG,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC;QAC3C,IAAI,CAAC,YAAY,GAAG,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC;QAC1C,IAAI,CAAC,cAAc,GAAG,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC;QAC5C,IAAI,CAAC,aAAa,GAAG,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC;QAE3C,wBAAwB;QACxB,MAAM,KAAK,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC;QACjC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7B,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,GAAG,KAAK,CAAC;YACtD,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,GAAG,KAAK,CAAC;QACvD,CAAC;IACH,CAAC;IAED;;OAEG;IACH,aAAa,CAAC,UAAsB;QAClC,IAAI,UAAU,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC;YAAE,OAAO;QAE1C,+BAA+B;QAC/B,MAAM,MAAM,GAAG,UAAU,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CACzC,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC,UAAU,CAAC,CACnC,CAAC;QAEF,+BAA+B;QAC/B,MAAM,UAAU,GAAG,IAAI,CAAC,UAAU,CAChC,UAAU,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC,EACnC,MAAM,CACP,CAAC;QAEF,kBAAkB;QAClB,MAAM,OAAO,GAAG,IAAI,CAAC,cAAc,CACjC,UAAU,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC,CACpC,CAAC;QAEF,gBAAgB;QAChB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,UAAU,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACjD,MAAM,IAAI,GAAG,UAAU,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;YACjC,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC;gBACf,KAAK,EAAE,IAAI,CAAC,UAAU;gBACtB,MAAM,EAAE,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC;gBACpC,MAAM,EAAE,IAAI,CAAC,MAAM;gBACnB,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC;gBAChB,OAAO,EAAE,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,MAAM,CAAC;gBAC1D,SAAS,EAAE,UAAU,CAAC,CAAC,CAAC;gBACxB,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC;aACpB,CAAC,CAAC;QACL,CAAC;IACH,CAAC;IAED;;;OAGG;IACH,MAAM;QACJ,MAAM,SAAS,GAAG,WAAW,CAAC,GAAG,EAAE,CAAC;QAEpC,IAAI,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,aAAa,EAAE,CAAC;YACnD,OAAO,EAAE,UAAU,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,CAAC;QACrD,CAAC;QAED,uBAAuB;QACvB,MAAM,UAAU,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC;QACrD,MAAM,OAAO,GAAG,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,GAAG,UAAU,CAAC,MAAM,CAAC;QAC1E,MAAM,MAAM,GAAG,IAAI,CAAC,IAAI,CACtB,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,GAAG,UAAU,CAAC,MAAM,CAC3E,GAAG,IAAI,CAAC;QAET,KAAK,MAAM,GAAG,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;YAC9B,GAAG,CAAC,SAAS,GAAG,CAAC,GAAG,CAAC,SAAS,GAAG,OAAO,CAAC,GAAG,MAAM,CAAC;QACrD,CAAC;QAED,IAAI,eAAe,GAAG,CAAC,CAAC;QACxB,IAAI,cAAc,GAAG,CAAC,CAAC;QACvB,IAAI,YAAY,GAAG,CAAC,CAAC;QACrB,IAAI,aAAa,GAAG,CAAC,CAAC;QACtB,IAAI,OAAO,GAAG,CAAC,CAAC;QAChB,IAAI,UAAU,GAAG,CAAC,CAAC;QAEnB,kBAAkB;QAClB,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,MAAM,EAAE,KAAK,EAAE,EAAE,CAAC;YACxD,iBAAiB;YACjB,IAAI,CAAC,aAAa,EAAE,CAAC;YAErB,uBAAuB;YACvB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,MAAM,EAAE,CAAC,IAAI,IAAI,CAAC,MAAM,CAAC,aAAa,EAAE,CAAC;gBACvE,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,CAAC;gBAClE,IAAI,KAAK,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,aAAa,GAAG,CAAC;oBAAE,SAAS;gBAE3D,MAAM,MAAM,GAAG,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,CAAC;gBAC3C,eAAe,IAAI,MAAM,CAAC,UAAU,CAAC;gBACrC,cAAc,IAAI,MAAM,CAAC,SAAS,CAAC;gBACnC,YAAY,IAAI,MAAM,CAAC,OAAO,CAAC;gBAC/B,aAAa,IAAI,MAAM,CAAC,QAAQ,CAAC;gBACjC,OAAO,IAAI,MAAM,CAAC,EAAE,CAAC;gBACrB,UAAU,EAAE,CAAC;gBAEb,gCAAgC;gBAChC,IAAI,MAAM,CAAC,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC,QAAQ,GAAG,GAAG,EAAE,CAAC;oBAC3C,MAAM;gBACR,CAAC;YACH,CAAC;QACH,CAAC;QAED,eAAe;QACf,IAAI,CAAC,MAAM,GAAG,EAAE,CAAC;QACjB,IAAI,CAAC,WAAW,EAAE,CAAC;QAEnB,MAAM,OAAO,GAAG,WAAW,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;QAC9C,IAAI,OAAO,GAAG,EAAE,EAAE,CAAC;YACjB,OAAO,CAAC,IAAI,CAAC,+BAA+B,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC;QAC7E,CAAC;QAED,OAAO;YACL,UAAU,EAAE,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC,eAAe,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC;YAC7D,SAAS,EAAE,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC,cAAc,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC;YAC3D,OAAO,EAAE,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC,YAAY,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC;SACxD,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,SAAS,CAAC,KAAmB;QAC3B,MAAM,MAAM,GAAG,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC;QACzC,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;QACnC,MAAM,MAAM,GAAG,IAAI,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC;QAExC,OAAO;YACL,MAAM;YACN,OAAO,EAAE,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,IAAI,CAAC;YACvC,KAAK,EAAE,IAAI,CAAC,YAAY,CAAC,KAAK,CAAC;SAChC,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,QAAQ;QACN,OAAO;YACL,WAAW,EAAE,IAAI,CAAC,WAAW;YAC7B,UAAU,EAAE,IAAI,CAAC,MAAM,CAAC,MAAM;YAC9B,OAAO,EAAE,IAAI,CAAC,WAAW,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC;YACrE,QAAQ,EAAE,IAAI,CAAC,QAAQ;YACvB,YAAY,EAAE,IAAI,CAAC,YAAY;SAChC,CAAC;IACJ,CAAC;IAED,6EAA6E;IAC7E,kBAAkB;IAClB,6EAA6E;IAErE,YAAY,CAAC,KAAmB;QACtC,IAAI,KAAK,GAAG,CAAC,CAAC;QACd,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,EAAE,IAAI,CAAC,YAAY,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC1E,KAAK,IAAI,KAAK,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC;QAC3C,CAAC;QACD,OAAO,KAAK,CAAC;IACf,CAAC;IAEO,aAAa,CAAC,KAAmB;QACvC,iCAAiC;QACjC,MAAM,UAAU,GAAG,CAAC,CAAC;QACrB,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,UAAU,CAAC,CAAC;QAE5C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,UAAU,EAAE,CAAC,EAAE,EAAE,CAAC;YACpC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,EAAE,IAAI,CAAC,aAAa,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC3E,MAAM,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,GAAG,CAAC,CAAC;YAChE,CAAC;QACH,CAAC;QAED,OAAO,MAAM,CAAC;IAChB,CAAC;IAEO,cAAc,CAAC,KAAmB,EAAE,MAAc;QACxD,MAAM,MAAM,GAAG,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC;QACzC,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;QACnC,MAAM,SAAS,GAAG,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;QAC1C,OAAO,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,SAAS,CAAC,GAAG,IAAI,CAAC,CAAC;IAC3C,CAAC;IAEO,UAAU,CAAC,MAAc;QAC/B,oCAAoC;QACpC,IAAI,IAAI,GAAG,CAAC,CAAC;QACb,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACvC,IAAI,GAAG,CAAC,IAAI,GAAG,EAAE,GAAG,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;QAChD,CAAC;QACD,OAAO,IAAI,CAAC;IACd,CAAC;IAEO,OAAO,CAAC,MAAoB;QAClC,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC;QAChC,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;QAC7C,IAAI,GAAG,GAAG,CAAC,CAAC;QAEZ,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACvC,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC;YACpC,GAAG,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC;QACjB,CAAC;QAED,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACrC,IAAI,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC;QACjB,CAAC;QAED,OAAO,IAAI,CAAC;IACd,CAAC;IAEO,YAAY,CAAC,KAAmB;QACtC,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC;QACxB,IAAI,MAAM,GAAG,CAAC,CAAC;QACf,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACtC,MAAM,IAAI,KAAK,CAAC,CAAC,CAAC,CAAC;YACnB,IAAI,CAAC,GAAG,MAAM;gBAAE,OAAO,CAAC,CAAC;QAC3B,CAAC;QACD,OAAO,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC;IAC1B,CAAC;IAEO,UAAU,CAAC,OAAiB,EAAE,MAAgB;QACpD,MAAM,UAAU,GAAG,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACrD,IAAI,OAAO,GAAG,CAAC,CAAC;QAEhB,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7C,MAAM,SAAS,GAAG,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAC7D,MAAM,KAAK,GAAG,OAAO,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,GAAG,SAAS,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;YACrE,OAAO,GAAG,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,SAAS,GAAG,OAAO,CAAC;YACtE,UAAU,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC;QAC1B,CAAC;QAED,OAAO,UAAU,CAAC;IACpB,CAAC;IAEO,cAAc,CAAC,OAAiB;QACtC,MAAM,OAAO,GAAG,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAClD,IAAI,SAAS,GAAG,CAAC,CAAC;QAElB,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7C,SAAS,GAAG,OAAO,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,GAAG,SAAS,CAAC;YACvD,OAAO,CAAC,CAAC,CAAC,GAAG,SAAS,CAAC;QACzB,CAAC;QAED,OAAO,OAAO,CAAC;IACjB,CAAC;IAEO,aAAa;QACnB,KAAK,IAAI,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAChD,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;YAC9C,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;QACtE,CAAC;IACH,CAAC;IAEO,eAAe,CAAC,KAAsB;QAO5C,IAAI,UAAU,GAAG,CAAC,CAAC;QACnB,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,IAAI,OAAO,GAAG,CAAC,CAAC;QAChB,IAAI,QAAQ,GAAG,CAAC,CAAC;QACjB,IAAI,EAAE,GAAG,CAAC,CAAC;QAEX,MAAM,UAAU,GAAG,IAAI,YAAY,CAAC,IAAI,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC;QAC/D,MAAM,SAAS,GAAG,IAAI,YAAY,CAAC,IAAI,CAAC,YAAY,CAAC,MAAM,CAAC,CAAC;QAE7D,KAAK,MAAM,GAAG,IAAI,KAAK,EAAE,CAAC;YACxB,iBAAiB;YACjB,MAAM,MAAM,GAAG,IAAI,CAAC,aAAa,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;YAC7C,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;YACnC,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC;YACtD,MAAM,YAAY,GAAG,IAAI,CAAC,YAAY,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;YAElD,gBAAgB;YAChB,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,GAAG,GAAG,CAAC,OAAO,CAAC,CAAC;YAEjD,8BAA8B;YAC9B,MAAM,KAAK,GAAG,KAAK,GAAG,GAAG,CAAC,SAAS,CAAC;YACpC,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CACpB,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,EAC1C,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,SAAS,CAC1B,GAAG,GAAG,CAAC,SAAS,CAAC;YAElB,MAAM,WAAW,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;YAC5C,UAAU,IAAI,WAAW,CAAC;YAE1B,iBAAiB;YACjB,IAAI,IAAI,CAAC,GAAG,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,SAAS,EAAE,CAAC;gBAChD,QAAQ,EAAE,CAAC;YACb,CAAC;YAED,8BAA8B;YAC9B,EAAE,IAAI,CAAC,GAAG,CAAC,OAAO,GAAG,UAAU,CAAC,CAAC;YAEjC,aAAa;YACb,IAAI,UAAkB,CAAC;YACvB,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW,KAAK,IAAI,EAAE,CAAC;gBACrC,MAAM,SAAS,GAAG,YAAY,CAAC;gBAC/B,MAAM,YAAY,GAAG,GAAG,CAAC,KAAK,GAAG,IAAI,CAAC,GAAG,CACvC,IAAI,CAAC,GAAG,CAAC,SAAS,GAAG,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,EACxD,CAAC,IAAI,CAAC,MAAM,CAAC,WAAW,CACzB,CAAC;gBACF,MAAM,GAAG,GAAG,CAAC,SAAS,GAAG,GAAG,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;gBAC3C,MAAM,GAAG,GAAG,CAAC,YAAY,GAAG,GAAG,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;gBAC9C,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC;YAClC,CAAC;iBAAM,CAAC;gBACN,UAAU,GAAG,CAAC,YAAY,GAAG,GAAG,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;YACjD,CAAC;YACD,SAAS,IAAI,UAAU,CAAC;YAExB,UAAU;YACV,IAAI,QAAQ,GAAG,CAAC,CAAC;YACjB,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;gBACtB,IAAI,CAAC,GAAG,CAAC;oBAAE,QAAQ,IAAI,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;YACzC,CAAC;YACD,OAAO,IAAI,QAAQ,CAAC;YAEpB,iCAAiC;YACjC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,EAAE,UAAU,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBACvE,UAAU,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,WAAW,GAAG,IAAI,CAAC;gBACnD,SAAS,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,UAAU,GAAG,IAAI,CAAC;YACnD,CAAC;QACH,CAAC;QAED,gCAAgC;QAChC,MAAM,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC,YAAY,CAAC;QACpC,MAAM,IAAI,GAAG,GAAG,CAAC;QAEjB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,aAAa,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACnD,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,GAAG,IAAI,GAAG,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC;YACpF,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,IAAI,EAAE,GAAG,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,CAAC;YAErD,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,GAAG,IAAI,GAAG,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC;YACjF,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,IAAI,EAAE,GAAG,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC;QACrD,CAAC;QAED,OAAO;YACL,UAAU,EAAE,UAAU,GAAG,KAAK,CAAC,MAAM;YACrC,SAAS,EAAE,SAAS,GAAG,KAAK,CAAC,MAAM;YACnC,OAAO,EAAE,OAAO,GAAG,KAAK,CAAC,MAAM;YAC/B,QAAQ,EAAE,QAAQ,GAAG,KAAK,CAAC,MAAM;YACjC,EAAE,EAAE,EAAE,GAAG,KAAK,CAAC,MAAM;SACtB,CAAC;IACJ,CAAC;CACF;AAED;;GAEG;AACH,MAAM,UAAU,SAAS,CAAC,MAA2B;IACnD,OAAO,IAAI,YAAY,CAAC,MAAM,CAAC,CAAC;AAClC,CAAC"}

package/dist/algorithms/q-learning.d.ts

@@ -0,0 +1,77 @@
+/**
+ * Tabular Q-Learning
+ *
+ * Classic Q-learning algorithm with:
+ * - Epsilon-greedy exploration
+ * - State hashing for continuous states
+ * - Eligibility traces (optional)
+ * - Experience replay
+ *
+ * Suitable for smaller state spaces or discretized environments.
+ * Performance Target: <1ms per update
+ */
+import type { Trajectory, RLConfig } from '../types.js';
+/**
+ * Q-Learning configuration
+ */
+export interface QLearningConfig extends RLConfig {
+    algorithm: 'q-learning';
+    explorationInitial: number;
+    explorationFinal: number;
+    explorationDecay: number;
+    maxStates: number;
+    useEligibilityTraces: boolean;
+    traceDecay: number;
+}
+/**
+ * Default Q-Learning configuration
+ */
+export declare const DEFAULT_QLEARNING_CONFIG: QLearningConfig;
+/**
+ * Q-Learning Algorithm Implementation
+ */
+export declare class QLearning {
+    private config;
+    private qTable;
+    private epsilon;
+    private stepCount;
+    private numActions;
+    private traces;
+    private updateCount;
+    private avgTDError;
+    constructor(config?: Partial<QLearningConfig>);
+    /**
+     * Update Q-values from trajectory
+     */
+    update(trajectory: Trajectory): {
+        tdError: number;
+    };
+    /**
+     * Get action using epsilon-greedy policy
+     */
+    getAction(state: Float32Array, explore?: boolean): number;
+    /**
+     * Get Q-values for a state
+     */
+    getQValues(state: Float32Array): Float32Array;
+    /**
+     * Get statistics
+     */
+    getStats(): Record<string, number>;
+    /**
+     * Reset Q-table
+     */
+    reset(): void;
+    private hashState;
+    private hashAction;
+    private getOrCreateEntry;
+    private updateTrace;
+    private updateWithTraces;
+    private pruneQTable;
+    private argmax;
+}
+/**
+ * Factory function
+ */
+export declare function createQLearning(config?: Partial<QLearningConfig>): QLearning;
+//# sourceMappingURL=q-learning.d.ts.map

package/dist/algorithms/q-learning.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"q-learning.d.ts","sourceRoot":"","sources":["../../src/algorithms/q-learning.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,KAAK,EAAE,UAAU,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAExD;;GAEG;AACH,MAAM,WAAW,eAAgB,SAAQ,QAAQ;IAC/C,SAAS,EAAE,YAAY,CAAC;IACxB,kBAAkB,EAAE,MAAM,CAAC;IAC3B,gBAAgB,EAAE,MAAM,CAAC;IACzB,gBAAgB,EAAE,MAAM,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,oBAAoB,EAAE,OAAO,CAAC;IAC9B,UAAU,EAAE,MAAM,CAAC;CACpB;AAED;;GAEG;AACH,eAAO,MAAM,wBAAwB,EAAE,eAetC,CAAC;AAWF;;GAEG;AACH,qBAAa,SAAS;IACpB,OAAO,CAAC,MAAM,CAAkB;IAGhC,OAAO,CAAC,MAAM,CAAkC;IAGhD,OAAO,CAAC,OAAO,CAAS;IACxB,OAAO,CAAC,SAAS,CAAK;IAGtB,OAAO,CAAC,UAAU,CAAK;IAGvB,OAAO,CAAC,MAAM,CAAwC;IAGtD,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,UAAU,CAAK;gBAEX,MAAM,GAAE,OAAO,CAAC,eAAe,CAAM;IAKjD;;OAEG;IACH,MAAM,CAAC,UAAU,EAAE,UAAU,GAAG;QAAE,OAAO,EAAE,MAAM,CAAA;KAAE;IA8EnD;;OAEG;IACH,SAAS,CAAC,KAAK,EAAE,YAAY,EAAE,OAAO,GAAE,OAAc,GAAG,MAAM;IAe/D;;OAEG;IACH,UAAU,CAAC,KAAK,EAAE,YAAY,GAAG,YAAY;IAW7C;;OAEG;IACH,QAAQ,IAAI,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC;IAUlC;;OAEG;IACH,KAAK,IAAI,IAAI;IAab,OAAO,CAAC,SAAS;IAejB,OAAO,CAAC,UAAU;IAQlB,OAAO,CAAC,gBAAgB;IAexB,OAAO,CAAC,WAAW;IAuBnB,OAAO,CAAC,gBAAgB;IAexB,OAAO,CAAC,WAAW;IAWnB,OAAO,CAAC,MAAM;CAWf;AAED;;GAEG;AACH,wBAAgB,eAAe,CAAC,MAAM,CAAC,EAAE,OAAO,CAAC,eAAe,CAAC,GAAG,SAAS,CAE5E"}

package/dist/algorithms/q-learning.js

@@ -0,0 +1,259 @@
+/**
+ * Tabular Q-Learning
+ *
+ * Classic Q-learning algorithm with:
+ * - Epsilon-greedy exploration
+ * - State hashing for continuous states
+ * - Eligibility traces (optional)
+ * - Experience replay
+ *
+ * Suitable for smaller state spaces or discretized environments.
+ * Performance Target: <1ms per update
+ */
+/**
+ * Default Q-Learning configuration
+ */
+export const DEFAULT_QLEARNING_CONFIG = {
+    algorithm: 'q-learning',
+    learningRate: 0.1,
+    gamma: 0.99,
+    entropyCoef: 0,
+    valueLossCoef: 1,
+    maxGradNorm: 1,
+    epochs: 1,
+    miniBatchSize: 1,
+    explorationInitial: 1.0,
+    explorationFinal: 0.01,
+    explorationDecay: 10000,
+    maxStates: 10000,
+    useEligibilityTraces: false,
+    traceDecay: 0.9,
+};
+/**
+ * Q-Learning Algorithm Implementation
+ */
+export class QLearning {
+    config;
+    // Q-table
+    qTable = new Map();
+    // Exploration
+    epsilon;
+    stepCount = 0;
+    // Number of actions
+    numActions = 4;
+    // Eligibility traces
+    traces = new Map();
+    // Statistics
+    updateCount = 0;
+    avgTDError = 0;
+    constructor(config = {}) {
+        this.config = { ...DEFAULT_QLEARNING_CONFIG, ...config };
+        this.epsilon = this.config.explorationInitial;
+    }
+    /**
+     * Update Q-values from trajectory
+     */
+    update(trajectory) {
+        const startTime = performance.now();
+        if (trajectory.steps.length === 0) {
+            return { tdError: 0 };
+        }
+        let totalTDError = 0;
+        // Reset eligibility traces for new trajectory
+        if (this.config.useEligibilityTraces) {
+            this.traces.clear();
+        }
+        for (let i = 0; i < trajectory.steps.length; i++) {
+            const step = trajectory.steps[i];
+            const stateKey = this.hashState(step.stateBefore);
+            const action = this.hashAction(step.action);
+            // Get or create Q-entry
+            const qEntry = this.getOrCreateEntry(stateKey);
+            // Current Q-value
+            const currentQ = qEntry.qValues[action];
+            // Compute target Q-value
+            let targetQ;
+            if (i === trajectory.steps.length - 1) {
+                // Terminal state
+                targetQ = step.reward;
+            }
+            else {
+                const nextStateKey = this.hashState(step.stateAfter);
+                const nextEntry = this.getOrCreateEntry(nextStateKey);
+                const maxNextQ = Math.max(...nextEntry.qValues);
+                targetQ = step.reward + this.config.gamma * maxNextQ;
+            }
+            // TD error
+            const tdError = targetQ - currentQ;
+            totalTDError += Math.abs(tdError);
+            if (this.config.useEligibilityTraces) {
+                // Update eligibility trace
+                this.updateTrace(stateKey, action);
+                // Update all states with traces
+                this.updateWithTraces(tdError);
+            }
+            else {
+                // Simple Q-learning update
+                qEntry.qValues[action] += this.config.learningRate * tdError;
+                qEntry.visits++;
+                qEntry.lastUpdate = Date.now();
+            }
+        }
+        // Decay exploration
+        this.stepCount += trajectory.steps.length;
+        this.epsilon = Math.max(this.config.explorationFinal, this.config.explorationInitial - this.stepCount / this.config.explorationDecay);
+        // Prune Q-table if too large
+        if (this.qTable.size > this.config.maxStates) {
+            this.pruneQTable();
+        }
+        this.updateCount++;
+        this.avgTDError = totalTDError / trajectory.steps.length;
+        const elapsed = performance.now() - startTime;
+        if (elapsed > 1) {
+            console.warn(`Q-learning update exceeded target: ${elapsed.toFixed(2)}ms > 1ms`);
+        }
+        return { tdError: this.avgTDError };
+    }
+    /**
+     * Get action using epsilon-greedy policy
+     */
+    getAction(state, explore = true) {
+        if (explore && Math.random() < this.epsilon) {
+            return Math.floor(Math.random() * this.numActions);
+        }
+        const stateKey = this.hashState(state);
+        const entry = this.qTable.get(stateKey);
+        if (!entry) {
+            return Math.floor(Math.random() * this.numActions);
+        }
+        return this.argmax(entry.qValues);
+    }
+    /**
+     * Get Q-values for a state
+     */
+    getQValues(state) {
+        const stateKey = this.hashState(state);
+        const entry = this.qTable.get(stateKey);
+        if (!entry) {
+            return new Float32Array(this.numActions);
+        }
+        return new Float32Array(entry.qValues);
+    }
+    /**
+     * Get statistics
+     */
+    getStats() {
+        return {
+            updateCount: this.updateCount,
+            qTableSize: this.qTable.size,
+            epsilon: this.epsilon,
+            avgTDError: this.avgTDError,
+            stepCount: this.stepCount,
+        };
+    }
+    /**
+     * Reset Q-table
+     */
+    reset() {
+        this.qTable.clear();
+        this.traces.clear();
+        this.epsilon = this.config.explorationInitial;
+        this.stepCount = 0;
+        this.updateCount = 0;
+        this.avgTDError = 0;
+    }
+    // ==========================================================================
+    // Private Methods
+    // ==========================================================================
+    hashState(state) {
+        // Discretize state by binning values
+        const bins = 10;
+        const parts = [];
+        // Use first 8 dimensions for hashing
+        for (let i = 0; i < Math.min(8, state.length); i++) {
+            const normalized = (state[i] + 1) / 2; // Assume [-1, 1] range
+            const bin = Math.floor(Math.max(0, Math.min(bins - 1, normalized * bins)));
+            parts.push(bin);
+        }
+        return parts.join(',');
+    }
+    hashAction(action) {
+        let hash = 0;
+        for (let i = 0; i < action.length; i++) {
+            hash = (hash * 31 + action.charCodeAt(i)) % this.numActions;
+        }
+        return hash;
+    }
+    getOrCreateEntry(stateKey) {
+        let entry = this.qTable.get(stateKey);
+        if (!entry) {
+            entry = {
+                qValues: new Float32Array(this.numActions),
+                visits: 0,
+                lastUpdate: Date.now(),
+            };
+            this.qTable.set(stateKey, entry);
+        }
+        return entry;
+    }
+    updateTrace(stateKey, action) {
+        // Decay all existing traces
+        for (const [key, trace] of this.traces) {
+            for (let a = 0; a < this.numActions; a++) {
+                trace[a] *= this.config.gamma * this.config.traceDecay;
+            }
+            // Remove near-zero traces
+            const maxTrace = Math.max(...trace);
+            if (maxTrace < 0.001) {
+                this.traces.delete(key);
+            }
+        }
+        // Set trace for current state-action
+        let trace = this.traces.get(stateKey);
+        if (!trace) {
+            trace = new Float32Array(this.numActions);
+            this.traces.set(stateKey, trace);
+        }
+        trace[action] = 1.0;
+    }
+    updateWithTraces(tdError) {
+        const lr = this.config.learningRate;
+        for (const [stateKey, trace] of this.traces) {
+            const entry = this.qTable.get(stateKey);
+            if (entry) {
+                for (let a = 0; a < this.numActions; a++) {
+                    entry.qValues[a] += lr * tdError * trace[a];
+                }
+                entry.visits++;
+                entry.lastUpdate = Date.now();
+            }
+        }
+    }
+    pruneQTable() {
+        // Remove least recently used states
+        const entries = Array.from(this.qTable.entries())
+            .sort((a, b) => a[1].lastUpdate - b[1].lastUpdate);
+        const toRemove = entries.length - Math.floor(this.config.maxStates * 0.8);
+        for (let i = 0; i < toRemove; i++) {
+            this.qTable.delete(entries[i][0]);
+        }
+    }
+    argmax(values) {
+        let maxIdx = 0;
+        let maxVal = values[0];
+        for (let i = 1; i < values.length; i++) {
+            if (values[i] > maxVal) {
+                maxVal = values[i];
+                maxIdx = i;
+            }
+        }
+        return maxIdx;
+    }
+}
+/**
+ * Factory function
+ */
+export function createQLearning(config) {
+    return new QLearning(config);
+}
+//# sourceMappingURL=q-learning.js.map