Page 134 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Model Training and 1) Machine learning models for personalization and adaptivity:
Fine-Tuning
• Classification models (Random Forests) to categorize interaction
patterns or infer emotional states from interaction proxies.
• Regression models to predict engagement scores or learning
trajectory.
• Contextual bandit algorithms for optimizing the sequence of
quests and content presentation.
2) NLP for content processing and generation:

• Sentiment analysis on caregiver feedback to identify areas of
concern or success.
• Text simplification algorithms to adjust readability of instructions
and story elements.
3) Predictive analytics for engagement forecasting:

• Sequence modeling (ex: LSTMs, Transformers if data volume
allows, otherwise simpler Markov models) on interaction logs to
predict when a child might become disengaged or frustrated.

4) Reinforcement learning for reward optimization:
• Models (Q-learning or policy gradient methods like PPO, with
appropriate simplifications for the domain) will be trained to opti-
mize the delivery (timing, type, magnitude) of rewards. The reward
function for the RL agent will be a composite function aiming to
maximize sustained engagement and learning progress, poten-
tially including terms for:
• + (Quest Completion Success)
• + (Correct Application of Moral in Scenario)
• – (Hints Used Penalty)
• + (Normalized Engagement Time within Optimal Range)
• – (Signs of Frustration/Disengagement from interaction
patterns)
Testbeds or Pilot In development, the Minimum Viable Product (MVP) phase is currently
Deployments underway. The MVP will have the following features.
1. Basic gamified mobile interface with simple touch interactions
2. A limited library of traditional Indian moral tales converted into
interactive quests
3. Fundamental AI personalization using basic classification models
for behavioral patterns
4. Simple dynamic difficulty adjustment based on rule-based thresh-
olds (ex: X correct answers increases difficulty)
5. Basic reward system (points, badges) without advanced AI optimi-
zation
6. Essential data collection from user interactions
7. Basic caregiver dashboard with simple progress metrics
8. Multi-modal content delivery (visual illustrations, audio narration,
text)
Code repositories Not publicly disclosed

98