Which robot actions actually caused the improvement?
An interactive demo showing how Shapley values identify which specific actions in a social robot intervention caused the observed outcome—and why RL temporal decay methods get this wrong.
Social robot interventions for children with ASD involve many actions over many days. When we observe improvement, the question is:
"Which of those N actions across X days actually caused the Y% gain?"
RL temporal decay methods (GAE/TD) assign credit based on recency—recent actions get most credit. This is systematically wrong: the pivotal action might have been early encouragement after a failed attempt, not the final session.
Shapley values treat each action as a "player" and compute its fair causal contribution by averaging marginal impact across all possible action orderings:
φᵢ = Σ [|S|!(n-|S|-1)!/n!] × [v(S ∪ {i}) - v(S)]
Key insight: Shapley values identify causal impact regardless of timing.
The demo shows a 15-day intervention with 35% improvement in emotion recognition:
| Day | Action | Shapley | RL Credit | Note |
|---|---|---|---|---|
| 3 | Encouragement | +12% | 2% | Turning point after failed attempt |
| 6 | Caregiver Prompt | +8% | 3% | Triadic interaction synergy |
| 15 | Difficulty Up | -2% | 21% | Counterproductive final push |
The contrast:
- Shapley correctly identifies Day 3 as most impactful (+12%)
- RL assigns Day 15 as most impactful (21%)—exactly wrong
| Action | Description |
|---|---|
| Emotion Game | Emotion recognition practice (Story game) |
| Perspective Game | Perspective-taking practice (Rocket/House game) |
| Sequencing Game | Sequencing practice (Train game) |
| Encouragement | Robot provides social encouragement |
| Difficulty ↑/↓ | BKT-driven difficulty adjustment |
| Caregiver Prompt | Robot prompts caregiver involvement |
- Salomons et al. (2018). "Robots for autism." Science Robotics.
- Shapley, L.S. (1953). "A Value for n-Person Games."
npm install
npm run dev # Development server
npm run build # Production build
npm run deploy # Deploy to GitHub PagesMIT License