 |
|
|
 |
|
|
|
|
|
|
MotivationSecurity at major locations of economic or political importance is a key concern around the world, particularly given the threat of terrorism. Unfortunately, in many applications, agents face human adversaries who --- because of their cognitive limitations --- may deviate from expected rational behavior. Not considering these likely deviations when dealing with human adversaries may cause an unacceptable degredation in security measures designed to defend against serious attacks. We work towards understanding these cognitive limitations through experimentation and develop new algorithms to help handle the deviations these limitations create. Description
Security, commonly defined as the ability to deal with intentional threats from
other agents, is a major challenge for agents or agent-teams deployed in
adversarial domains. Such adversarial scenarios arise in a wide variety of
situations that are becoming increasingly important. Some example cases are
agents patrolling to provide perimeter security around critical infrastructure
or performing routine security checks. Such security domains are often modeled as a Stackelberg game, where the defender (leader) commits to a strategy first and the adversary (follower) observers such strategy and response to it. Classical game theory provides a nice framework to solve such game for the defender based on a set of assumptions of adversaries' decision-making, such as the adversaries are perfect rational and they have complete information of the game. However, in many real security domains, defender is facing real human adversaries and most of these assumptions may not hold. Therefore, it is very important to generate more robust strategy for the defender by integrating more realistic model of human decision-making. Anchoring Bias is when, given no information about the occurrence of a discrete set of events, humans will tend to assign an equal weight to the occurrence of each event (a uniform distribution). Bounded Rationality: the follower is boundedly rational and may not strictly maximize utility. As a result, the follower may select an epslion-optimal response strategy, i.e. the follower may choose any of the responses within epsilon-reward of her optimal strategy. Prospect Theory is a noble-prize-winning theory, which describes human decision making as a process of maximizing `prospect'. Quantal Response: Quantal Response Equilibrium suggests that instead of strictly maximizing utility, individuals respond stochastically in games: the chance of selecting a non-optimal strategy increases as the cost of such an error decreases. We have developed COBRA which accounts for the anchoring bias human adversaries may experience when interpreting the probability of defender's mixed-strategy and their bounded rationality in computing the best response. In the experiment with human subjects, COBRA outperforms the perfect rational model (DOBSS) with statistical significance. We also work toward developing new algorithms to account for the cognitive limitations of human adversaries. We have applied two fundamental theories of human behavior to predict an attacker's decision: Prospect Theory and Quantal Response Equilibrium. In the experiments with human subjects, BRQR (i.e. algorithm to compute best response of defender against quantal response of adversary) outperforms COBRA with statistical significance.
The Game: We developed the online game [The Guards and The Treasure] to simulate security senario at the LAX airport. Play the game, help us understand how human behavior in the security game. We are trying to design better strategy to protect the facilities against real attackers. Recently, we have developed three new algorithms which are based on two important models in human behavior literature. We tested our new algorithms using the game with human subjects. Our new algorithms outperformed the current leading contender with statistical sigfinicance. In total, we we tested 7 different payoff structures. For each payoff structures, we generated defender strategies for the following five different algorithms to play against human subjects. The score (average expected utility for the defender) of each stragies are displayed in the following graph for each payoff structure.
Overall, BRQR showed best performance; DOBSS and BRPT had the worst performances. Relevant Papers and Presentations
If you have any questions about the contents of this page please contact James Pita ( jpita@usc.edu ) |
