8 Counterfactuals and parallel worlds
This chapter covers
- Motivating examples for counterfactual reasoning
- Turning counterfactual questions into symbolic form
- Building parallel world graphs for counterfactual reasoning
- Implementing the counterfactual inference algorithm
- Building counterfactual deep generative models of images
Marjani, a good friend of mine, once had to choose between two dating prospects at the same time. She had something of a mental score card for an ideal long-term match. She had good chemistry with one guy, but he didn’t rank well on the score card. In contrast, the second guy checked all the boxes, so she chose him. But after some time, despite him meeting all her criteria, she couldn’t muster any feelings for him. It was like a failed ritual summoning; the stars were perfectly aligned, but the summoned spirit never showed up. And so, as any of us would in that situation, she posed the counterfactual question:
I chose a partner based on my criteria and it’s not working out. Would it have worked out if I chose based on chemistry?
Counterfactual queries like this describe hypothetical events that did not occur but could have occurred if something had been different. Counterfactuals are fundamental to how we define causality; if the answer to Marjani’s question is yes, it implies that choosing based on her score card caused her love life to be unsuccessful.