学术文献库

We seek to understand the fundamental mathematics governing infrastructure-scale interactions between humans and machines, particularly when the machines' intended purpose is to influence and optimize the behavior of the humans. To that end, this paper investigates the worst-case congestion that can arise in nonatomic network congestion games when a fraction of the traffic is completely altruistic (e.g., benevolent self-driving cars) and the remainder is completely selfish (e.g., human commuters). We study the worst-case harm of altruism in such scenarios in terms of the perversity index, or the worst-case equilibrium congestion cost resulting from the presence of altruistic traffic, relative to the congestion cost which would result if all traffic were selfish. We derive a tight bound on the perversity index for the class of series-parallel network congestion games with convex latency functions, and show three facts: First, the harm of altruism is maximized when exactly half of the traffic is altruistic, but it gracefully vanishes when the fraction of altruistic traffic approaches either 0 or 1. Second, we show that the harm of altruism is linearly increasing in a natural measure of the "steepness" of network latency functions. Finally, we show that for any nontrivial fraction of altruistic traffic, the harm of altruism exceeds the price of anarchy associated with all-selfish traffic: in a sense, partial altruism is worse than complete selfishness.