Cathy l posted an update — Jan 16, 2026 01:01 PM EST

Reinforcement Learning System

Reinforcement learns from user feedback and outcome data:

3 Main Components:

1. core/reinforcement.py - Core RL System

   • ReinforcementScorer: Calculates reward from YouTube metrics
   • DecisionRewardTracker: Tracks decisions + outcomes
   • UserFeedbackLearner: Learns from user corrections
   • AdaptiveDecisionMaker: Makes decisions based on learned patterns
   • LearningAgent: Orchestrates the learning loop

2. core/integrated_decision_engine.py - Updated Decision Engine

   • Integrates Gemini reasoning with learned patterns
   • Records decisions for learning
   • Improves decisions based on historical performance
   • Respects user overrides first

3. scripts/update_learning.py - Feedback Collection

   • Fetches YouTube performance data
   • Records user feedback & corrections
   • Analyzes learning progress
   • Usage: python scripts/update_learning.py --fetch-youtube

4. scripts/learning_dashboard.py - Visualization

   • Interactive dashboard showing:
     • Decision quality trends
     • User feedback patterns
     • Learned preferences
     • Performance improvements
   • Multiple views: main, history, feedback, trends

How It Works:

1. Record Decision → Agent makes decision with reasoning
2. Get Outcome → YouTube performance + user feedback
3. Calculate Reward → Score based on views, likes, engagement
4. Learn → Update patterns from reward
5. Improve → Next similar decision uses learned data

User Feedback Loop:

• User changes privacy: "private" → "public" = -0.5 reward
• User keeps decision = +reinforcement
• User rates 5 stars = +0.8 reward
• Patterns accumulated → Inferred preferences

Result: Agent gets smarter with every video, learning your actual preferences without retraining models!

Log in or sign up for Devpost to join the conversation.