Planning with Multiple Biases

Jon M. Kleinberg, Sigal Oren, Manish Raghavan

Abstract

While behavioral economics has long studied individual cognitive biases, real-world planning often involves multiple biases acting in concert. This paper introduces a theoretical framework to model the interaction between present bias (the urge to procrastinate) and sunk-cost bias (the tendency to "throw good money after bad"). Our analysis reveals that these biases can either mitigate or catastrophically amplify each other, sometimes making the simple act of planning computationally NP-hard.

Introduction

Traditional economic models assume "rational agents" who optimize for maximum utility. Behavioral economics counters this by identifying specific biases, such as present bias (overvaluing immediate costs/benefits) and sunk-cost bias (incorporating unrecoverable past costs into future decisions). However, most research treats these biases in isolation. In reality, agents often experience both simultaneously. For instance, an individual might pay for an expensive gym membership (sunk cost) to motivate their future self (present bias) to exercise. This paper explores the formal interaction between these biases, asking whether they cancel each other out or lead to even more sub-optimal behavior.

The Graph-Theoretic Model

The authors utilize a graph-theoretic model to represent the planning process. An agent moves through a directed graph from a start node $s$ to a target node $t$ to claim a reward $R$. Each edge in the graph represents a task with an associated cost.

Figure 1: A basic graph-theoretic model for planning. An agent chooses a path from $s$ to $t$ based on perceived costs and a potential reward $R$.

The model incorporates two primary biases:

  • Present Bias ($b$): When at a node, the agent perceives the cost of the immediate next step as being $b$ times larger than its actual value ($b \ge 1$).
  • Sunk-Cost Bias ($\lambda$): If an agent considers quitting, they experience a "mental cost" proportional to the total cost already incurred ($\lambda \times \text{sunk cost}$).

The paper further distinguishes between naive agents, who are unaware of their biases, and sophisticated agents, who accurately predict their future biased behavior and plan accordingly.

Key Insights: Synergy and Complexity

The interaction between these biases leads to several counterintuitive behaviors.

Sunk Cost as a Commitment Device

A sophisticated agent can strategically use their sunk-cost bias to overcome present bias. By incurring a high cost early in a project, the agent creates a high "mental penalty" for quitting, effectively forcing their future, procrastinating self to complete the task.

The Complexity Explosion

One of the most significant findings is that planning for a "doubly sophisticated" agent (aware of both biases) is NP-hard. While planning with a single bias is typically computationally straightforward, the interplay between past expenditures and future perceptions creates dependencies that mirror the Subset Sum problem.

Figure 2: A complex layered graph structure used to demonstrate the computational complexity and potential for exponential waste in multi-bias planning.

Exponential Waste

For "doubly naive" agents, certain graph structures (like the "fan-like" structure shown in Figure 2) can lead to catastrophic outcomes. These agents may be lured into paths where they incur costs exponentially larger than the reward $R$, constantly believing they are near completion while the sunk-cost bias prevents them from abandoning the increasingly expensive endeavor.

Key Results

The researchers established several bounds and algorithmic results:

  • Payoff Gaps: For a doubly sophisticated agent, the difference between their payoff and the optimal rational payoff is bounded by $(b-1)C_o$, where $C_o$ is the optimal cost. This suggests that if they reach the goal, they do not perform significantly worse than a rational agent.
  • Pseudo-Polynomial Algorithm: A planning algorithm for sophisticated agents exists that runs in time polynomial to the number of nodes and the total sum of edge costs.
  • The "Singly Sophisticated" Trap: Agents aware of present bias but naive about sunk-cost bias remain time-inconsistent, often changing plans mid-journey because they fail to anticipate how past spending will influence their future motivation.

Conclusion and Future Work

This work demonstrates that the interaction of cognitive biases is not merely additive but can fundamentally change the computational and behavioral nature of planning. Future research should explore whether planning remains efficient when rewards are significantly larger than costs and identify specific graph properties that trigger "exponentially bad" behavior. Additionally, extending this model to include more than two biases remains a vital area for understanding human decision-making in complex environments.

Summary updated

BibTeX
@inproceedings{kleinberg2017planning,
  author = {Jon M. Kleinberg and Sigal Oren and Manish Raghavan},
  title = {{Planning with Multiple Biases}},
  booktitle = {EC},
  year = {2017},
  url = {https://openreview.net/forum?id=6fjw2U_997m},
}
Figure