סיגל אורן

People often believe that their future preferences will be similar to their current ones. For example, people who go hungry to the supermarket, often buy less healthy food items than when they go on a full stomach. Loewenstein et al. [10] coined the term projection bias to capture this and similar behaviors. Our first contribution is a generalization of the restricted model of Loewenstein et al. by considering agents with projection bias that traverse a state graph for time horizon t. Our generalization allows us to capture more complex planning scenarios, such as a student that plans his occupational path. We analyze the planning behavior of biased agents and show that their loss due to their projection bias may be unbounded. Obviously, agents who do not suffer from projection bias at all will be able to traverse the graph optimally. We show–perhaps surprisingly–that agents that exhibit a strong projection bias sometimes fare better than agents that exhibit projection bias to a smaller extent. Similarly, we show that agents that plan for a longer time horizon do not necessarily fare better than agents that plan for a shorter time horizon. We then provide bounds on the number of these “non-monotonicity” points in a given state graph. Among other results, we prove a hardness result for computing a subgraph that maximizes the utility of the biased agent.