agentdb 1.0.0

package/dist/plugins/implementations/curriculum-learning.mjs

@@ -0,0 +1,349 @@
+/**
+ * Curriculum Learning Plugin
+ *
+ * Implements structured learning where the model progressively tackles
+ * increasingly difficult tasks, mimicking human learning.
+ *
+ * Key features:
+ * - Automatic difficulty estimation
+ * - Dynamic curriculum generation
+ * - Self-paced learning
+ * - Task sequencing based on prerequisites
+ * - Competence-based progression
+ */
+import { BasePlugin } from '../base-plugin.mjs';
+/**
+ * Curriculum Learning Plugin Implementation
+ */
+export class CurriculumLearningPlugin extends BasePlugin {
+    constructor(config) {
+        super();
+        this.name = 'curriculum-learning';
+        this.version = '1.0.0';
+        this.experiences = [];
+        this.curriculum = [];
+        this.currentStage = 0;
+        this.taskDifficulties = new Map();
+        this.competence = new Map();
+        this.strategy = 'automatic';
+        // Learning parameters
+        this.minSuccessRate = 0.7;
+        this.competenceThreshold = 0.8;
+        this.difficultyWindow = 0.2;
+        if (config) {
+            this.strategy = config.strategy || 'automatic';
+            this.minSuccessRate = config.minSuccessRate || 0.7;
+            this.competenceThreshold = config.competenceThreshold || 0.8;
+        }
+        this.initializeCurriculum();
+        // Mark as initialized for in-memory operation
+        this.initialized = true;
+    }
+    /**
+     * Override to skip initialization check for in-memory operation
+     */
+    checkInitialized() {
+        // No-op for curriculum learning - operates in-memory
+    }
+    /**
+     * Override selectAction to provide base implementation
+     */
+    async selectAction(state, context) {
+        // Simple default action selection
+        const stateSum = state.reduce((a, b) => a + b, 0);
+        const value = Math.tanh(stateSum / state.length);
+        return {
+            id: String(Math.floor(Math.abs(value) * 10)),
+            embedding: state,
+            confidence: Math.abs(value),
+        };
+    }
+    /**
+     * Override to store experiences in-memory without vectorDB
+     */
+    async storeExperience(experience) {
+        this.experiences.push(experience);
+        const taskId = experience.metadata?.taskId || 'unknown';
+        const success = experience.reward > 0;
+        // Update competence
+        this.updateCompetence(taskId, experience.reward, success);
+        // Update task difficulty estimate using local experiences
+        const taskExps = this.experiences.filter(e => e.metadata?.taskId === taskId);
+        const difficulty = this.estimateTaskDifficulty(taskId, taskExps);
+        this.taskDifficulties.set(taskId, difficulty);
+    }
+    /**
+     * Override to retrieve from local experiences
+     */
+    async retrieveSimilar(state, k) {
+        return this.experiences.slice(0, k).map((exp, idx) => ({
+            id: exp.id || `exp-${idx}`,
+            embedding: exp.state,
+            metadata: exp,
+            score: 1.0 - (idx * 0.1),
+        }));
+    }
+    /**
+     * Initialize curriculum stages
+     */
+    initializeCurriculum() {
+        this.curriculum = [
+            {
+                stage: 0,
+                name: 'Foundation',
+                minCompetence: 0.0,
+                tasks: ['basic_actions', 'simple_rewards'],
+                completed: false,
+            },
+            {
+                stage: 1,
+                name: 'Intermediate',
+                minCompetence: 0.7,
+                tasks: ['sequential_actions', 'delayed_rewards'],
+                completed: false,
+            },
+            {
+                stage: 2,
+                name: 'Advanced',
+                minCompetence: 0.8,
+                tasks: ['complex_strategies', 'multi_step_planning'],
+                completed: false,
+            },
+            {
+                stage: 3,
+                name: 'Expert',
+                minCompetence: 0.9,
+                tasks: ['optimal_policies', 'transfer_learning'],
+                completed: false,
+            },
+        ];
+    }
+    /**
+     * Estimate task difficulty from experiences
+     */
+    estimateTaskDifficulty(taskId, experiences) {
+        if (experiences.length === 0) {
+            return {
+                taskId,
+                complexity: 0.5,
+                estimatedLoss: 1.0,
+                successRate: 0.0,
+                prerequisites: [],
+            };
+        }
+        // Calculate metrics
+        const successes = experiences.filter(e => e.reward > 0).length;
+        const successRate = successes / experiences.length;
+        const avgReward = experiences.reduce((sum, e) => sum + e.reward, 0) / experiences.length;
+        const rewardVariance = experiences.reduce((sum, e) => sum + Math.pow(e.reward - avgReward, 2), 0) / experiences.length;
+        // Complexity based on state dimensionality and reward variance
+        const stateDim = experiences[0].state.length;
+        const complexity = Math.min(1.0, (stateDim / 256 + Math.sqrt(rewardVariance)) / 2);
+        return {
+            taskId,
+            complexity,
+            estimatedLoss: 1.0 - successRate,
+            successRate,
+            prerequisites: this.inferPrerequisites(taskId, complexity),
+        };
+    }
+    /**
+     * Infer prerequisite tasks based on complexity
+     */
+    inferPrerequisites(taskId, complexity) {
+        const prerequisites = [];
+        this.taskDifficulties.forEach((difficulty, id) => {
+            if (id !== taskId && difficulty.complexity < complexity - 0.2) {
+                prerequisites.push(id);
+            }
+        });
+        return prerequisites.slice(0, 3); // Max 3 prerequisites
+    }
+    /**
+     * Select next task based on current competence
+     */
+    selectNextTask() {
+        const currentCompetence = this.getCurrentCompetence();
+        // Self-paced: Choose task slightly above current competence
+        if (this.strategy === 'self_paced') {
+            return this.selectSelfPacedTask(currentCompetence);
+        }
+        // Automatic: Estimate difficulty and choose optimally
+        if (this.strategy === 'automatic') {
+            return this.selectAutomaticTask(currentCompetence);
+        }
+        // Predefined: Follow curriculum stages
+        return this.selectCurriculumTask();
+    }
+    /**
+     * Self-paced task selection
+     */
+    selectSelfPacedTask(competence) {
+        const targetDifficulty = competence + this.difficultyWindow;
+        let bestTask = null;
+        let bestDiff = Infinity;
+        this.taskDifficulties.forEach((difficulty, taskId) => {
+            const diff = Math.abs(difficulty.complexity - targetDifficulty);
+            if (diff < bestDiff && this.arePrerequisitesMet(taskId)) {
+                bestTask = taskId;
+                bestDiff = diff;
+            }
+        });
+        return bestTask;
+    }
+    /**
+     * Automatic task selection using competence-based heuristic
+     */
+    selectAutomaticTask(competence) {
+        const candidates = [];
+        this.taskDifficulties.forEach((difficulty, taskId) => {
+            if (!this.arePrerequisitesMet(taskId)) {
+                return;
+            }
+            // Score combines difficulty match and learning progress potential
+            const difficultyMatch = 1.0 - Math.abs(difficulty.complexity - competence);
+            const learningPotential = 1.0 - difficulty.successRate;
+            const score = difficultyMatch * 0.7 + learningPotential * 0.3;
+            candidates.push({ taskId, score });
+        });
+        if (candidates.length === 0) {
+            return null;
+        }
+        // Select best scoring task
+        candidates.sort((a, b) => b.score - a.score);
+        return candidates[0].taskId;
+    }
+    /**
+     * Curriculum-based task selection
+     */
+    selectCurriculumTask() {
+        const stage = this.curriculum[this.currentStage];
+        if (!stage || stage.completed) {
+            // Move to next stage
+            if (this.currentStage < this.curriculum.length - 1) {
+                this.currentStage++;
+                return this.selectCurriculumTask();
+            }
+            return null;
+        }
+        // Select random task from current stage
+        const uncompletedTasks = stage.tasks.filter(taskId => (this.competence.get(taskId) || 0) < this.competenceThreshold);
+        if (uncompletedTasks.length === 0) {
+            stage.completed = true;
+            return this.selectCurriculumTask();
+        }
+        return uncompletedTasks[Math.floor(Math.random() * uncompletedTasks.length)];
+    }
+    /**
+     * Check if task prerequisites are met
+     */
+    arePrerequisitesMet(taskId) {
+        const difficulty = this.taskDifficulties.get(taskId);
+        if (!difficulty) {
+            return true;
+        }
+        return difficulty.prerequisites.every(prereqId => {
+            const competence = this.competence.get(prereqId) || 0;
+            return competence >= this.minSuccessRate;
+        });
+    }
+    /**
+     * Get current overall competence
+     */
+    getCurrentCompetence() {
+        if (this.competence.size === 0) {
+            return 0.0;
+        }
+        const sum = Array.from(this.competence.values()).reduce((a, b) => a + b, 0);
+        return sum / this.competence.size;
+    }
+    /**
+     * Update competence for a task
+     */
+    updateCompetence(taskId, reward, success) {
+        const currentCompetence = this.competence.get(taskId) || 0;
+        // Exponential moving average
+        const alpha = 0.1;
+        const newCompetence = success ? 1.0 : 0.0;
+        const updated = (1 - alpha) * currentCompetence + alpha * newCompetence;
+        this.competence.set(taskId, updated);
+    }
+    /**
+     * Train with curriculum learning
+     */
+    async train(options) {
+        const startTime = Date.now();
+        const epochs = options?.epochs || 50;
+        let totalLoss = 0;
+        let experiencesProcessed = 0;
+        for (let epoch = 0; epoch < epochs; epoch++) {
+            // Select next task based on curriculum
+            const taskId = this.selectNextTask();
+            if (!taskId) {
+                // Curriculum completed
+                break;
+            }
+            // Retrieve experiences for this task from local storage
+            const allExperiences = await this.retrieveSimilar(new Array(128).fill(0), options?.batchSize || 32);
+            const experiences = this.experiences
+                .filter(e => e.metadata?.taskId === taskId)
+                .slice(0, options?.batchSize || 32);
+            // Train on task
+            for (const exp of experiences) {
+                const prediction = await this.selectAction(exp.state);
+                const loss = Math.pow(prediction.confidence - exp.reward, 2);
+                totalLoss += loss;
+                experiencesProcessed++;
+            }
+            // Check if we should advance curriculum stage
+            if (this.strategy === 'predefined') {
+                const currentCompetence = this.getCurrentCompetence();
+                const nextStage = this.curriculum[this.currentStage + 1];
+                if (nextStage && currentCompetence >= nextStage.minCompetence) {
+                    this.currentStage++;
+                }
+            }
+        }
+        const duration = Math.max(1, Date.now() - startTime);
+        return {
+            loss: totalLoss / Math.max(1, experiencesProcessed),
+            experiencesProcessed,
+            duration,
+            currentStage: this.currentStage,
+            currentCompetence: this.getCurrentCompetence(),
+            tasksCompleted: Array.from(this.competence.entries())
+                .filter(([, comp]) => comp >= this.competenceThreshold)
+                .length,
+        };
+    }
+    /**
+     * Get curriculum progress
+     */
+    getCurriculumProgress() {
+        const totalTasks = this.curriculum.reduce((sum, stage) => sum + stage.tasks.length, 0);
+        const completed = Array.from(this.competence.values())
+            .filter(c => c >= this.competenceThreshold).length;
+        return {
+            currentStage: this.currentStage,
+            totalStages: this.curriculum.length,
+            competence: this.getCurrentCompetence(),
+            tasksCompleted: completed,
+            totalTasks,
+        };
+    }
+    /**
+     * Add custom curriculum stage
+     */
+    addCurriculumStage(stage) {
+        this.curriculum.push({
+            ...stage,
+            completed: false,
+        });
+    }
+    /**
+     * Get task difficulties
+     */
+    getTaskDifficulties() {
+        return new Map(this.taskDifficulties);
+    }
+}

package/dist/plugins/implementations/decision-transformer.d.ts

@@ -0,0 +1,77 @@
+/**
+ * Decision Transformer Plugin
+ *
+ * Implements the Decision Transformer algorithm with 3-tier action selection:
+ * - Tier 1: Exact retrieval (99% similarity)
+ * - Tier 2: k-NN interpolation (95% similarity)
+ * - Tier 3: Neural network generation
+ *
+ * Based on "Decision Transformer: Reinforcement Learning via Sequence Modeling"
+ * https://arxiv.org/abs/2106.01345
+ */
+import { BasePlugin } from '../base-plugin';
+import { Action, Context, TrainOptions, TrainingMetrics } from '../learning-plugin.interface';
+/**
+ * Decision Transformer Plugin Implementation
+ */
+export declare class DecisionTransformerPlugin extends BasePlugin {
+    name: string;
+    version: string;
+    private decisionLayer;
+    private tier1Threshold;
+    private tier2Threshold;
+    private kNeighbors;
+    /**
+     * Initialize the Decision Transformer plugin
+     */
+    protected onInitialize(): Promise<void>;
+    /**
+     * Select action using 3-tier strategy
+     *
+     * @param state - Current state vector
+     * @param context - Optional context including desired return
+     * @returns Selected action
+     */
+    selectAction(state: number[], context?: Context): Promise<Action>;
+    /**
+     * Tier 1: Try exact retrieval
+     */
+    private tryExactRetrieval;
+    /**
+     * Tier 2: Try k-NN interpolation
+     */
+    private tryKNNInterpolation;
+    /**
+     * Tier 3: Use neural network generation
+     */
+    private useNeuralGeneration;
+    /**
+     * Train the decision layer on stored experiences
+     */
+    train(options?: TrainOptions): Promise<TrainingMetrics>;
+    /**
+     * Compute return-to-go from experience metadata
+     */
+    private computeReturnToGo;
+    /**
+     * Encode history as average of recent states
+     */
+    private encodeHistory;
+    /**
+     * Get all experiences (helper method)
+     */
+    private getAllExperiences;
+    /**
+     * Group experiences by episode
+     */
+    private groupByEpisode;
+    /**
+     * Save model weights
+     */
+    protected onSave(path: string): Promise<void>;
+    /**
+     * Load model weights
+     */
+    protected onLoad(path: string): Promise<void>;
+}
+//# sourceMappingURL=decision-transformer.d.ts.map

package/dist/plugins/implementations/decision-transformer.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"decision-transformer.d.ts","sourceRoot":"","sources":["../../../src/plugins/implementations/decision-transformer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,gBAAgB,CAAC;AAC5C,OAAO,EACL,MAAM,EACN,OAAO,EAEP,YAAY,EACZ,eAAe,EAChB,MAAM,8BAA8B,CAAC;AA+OtC;;GAEG;AACH,qBAAa,yBAA0B,SAAQ,UAAU;IAChD,IAAI,SAA0B;IAC9B,OAAO,SAAW;IAEzB,OAAO,CAAC,aAAa,CAAiB;IACtC,OAAO,CAAC,cAAc,CAAgB;IACtC,OAAO,CAAC,cAAc,CAAgB;IACtC,OAAO,CAAC,UAAU,CAAa;IAE/B;;OAEG;cACa,YAAY,IAAI,OAAO,CAAC,IAAI,CAAC;IAa7C;;;;;;OAMG;IACG,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,EAAE,OAAO,CAAC,EAAE,OAAO,GAAG,OAAO,CAAC,MAAM,CAAC;IAgCvE;;OAEG;YACW,iBAAiB;IA2B/B;;OAEG;YACW,mBAAmB;IAqCjC;;OAEG;YACW,mBAAmB;IAsBjC;;OAEG;IACG,KAAK,CAAC,OAAO,CAAC,EAAE,YAAY,GAAG,OAAO,CAAC,eAAe,CAAC;IAiD7D;;OAEG;IACH,OAAO,CAAC,iBAAiB;IAIzB;;OAEG;IACH,OAAO,CAAC,aAAa;IAqBrB;;OAEG;YACW,iBAAiB;IAU/B;;OAEG;IACH,OAAO,CAAC,cAAc;IAqBtB;;OAEG;cACa,MAAM,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC;IAMnD;;OAEG;cACa,MAAM,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC;CAKpD"}