Skip to content

Backward Induction

Why It Matters

In sequential games, naive analysis focuses on immediate payoffs, ignoring how current choices constrain future options. Backward induction forces consideration of the full game tree: your current best move depends on what optimal play looks like after every possible action you might take. This reveals counter-intuitive strategies (sacrificing immediate gain for better position), exposes incredible threats (threats you wouldn't actually execute when the time comes), and enables optimal planning in multi-stage situations. Without backward induction, sequential reasoning devolves into myopic optimization.

The Key Move

  1. Map the sequential game as a decision tree with decision nodes (who chooses) and terminal nodes (outcomes).
  2. Start at the terminal nodes (end states) and note the payoffs.
  3. Move to the last decision nodes - the final choices before outcomes.
  4. For each player at these nodes, identify which action maximizes their payoff given the terminal payoffs.
  5. Replace those decision nodes with the payoff resulting from optimal choice.
  6. Move backward to the second-to-last decision nodes, repeat the optimization assuming optimal play afterward.
  7. Continue backward until reaching the initial node.
  8. The backward induction solution is the path of optimal choices from start to finish.

Classic Application

Centipede game. Two players alternate choosing "continue" or "stop." Pool of money grows each round. When you stop, you take slightly more than half, leaving the rest to the other player. Backward induction: at the final decision node, the last mover takes the full pot (stopping is optimal). Knowing this, the second-to-last mover should stop (since continuing gives the last mover the opportunity to take everything). Working backward, this logic unravels to the first decision: the first mover should stop immediately. Prediction: immediate stopping. Reality: humans often cooperate for several rounds, suggesting limits of pure backward induction with real players.

Surprising Application

Negotiation deadlines and timing. Two parties negotiating over a deal with a known deadline. Backward induction: work backward from the deadline. At the deadline moment, whoever needs the deal more has zero leverage (their alternative is no deal). Knowing this, the other party can extract maximum concession at the deadline. Knowing this, concessions should happen earlier to preserve some leverage. But working further backward, this logic suggests concessions should happen immediately. This explains why many negotiations have "deadline rush" - both parties avoid early concession (preserving leverage) but know that waiting until the deadline means capitulation. Non-obvious because it reveals how known endpoints propagate backward to affect early-stage behavior.

Failure Modes

Assuming perfect rationality throughout - backward induction requires all players to be perfectly rational at every stage, which humans aren't. Ignoring incredible threats - backward induction reveals that many threats won't be executed when the time comes, but humans sometimes irrationally follow through. Missing subgame imperfection - some equilibria involve non-credible strategies that backward induction eliminates. Not checking for multiple equilibria - some sequential games have multiple backward induction solutions. Applying to simultaneous games - backward induction requires sequential structure with observation.

Go Deeper

Osborne & Rubinstein, A Course in Game Theory, Chapter 6; Dixit & Skeath, Games of Strategy, Chapter 3