Based on observations of an opponent's past behavior, a new software system is trying to predict what the opponent will do next. DARPA, the US government's Defense Advanced Research Projects Agency, is funding the project in hopes that it may be useful in future battles.
Talk about having a psychological edge.
H. Van Dyke Parunak, a software engineer at a company called New Vectors, is leading the project, and recently submitted a patent application. Artificial intelligence (AI) experts have an interest in strategy-predicting software not only for military combat situations, but also for competitive business strategies and multi-player computer games.
In their attempts at this "mind reading," AI programs begin by describing a person's internal state: their beliefs, desires, intentions, and emotions. In the model developed by Parunak and his colleagues, a person's internal state is estimated by his behavior, such as his movement within the simulated environment.
Since a person's internal state-in a sense, his character-influences his actions, knowing that internal state could help the software make predictions about the person's actions in the future.
"An agent's goals guide its actions," the inventors explained in their patent application. "Thus one ought to be able to learn something about an agent's goals by observing its past actions, and knowledge of the agent's goals in turn enables conclusions about what the agent may do in the future."
In Parunak's simulation, future predictions can be made in complex environments, such as real-life situations. By "evolving" agents in a complex simulated environment, the researchers can observe their behavior, actions and reactions to events, and find a best-fit internal state based on these actions.
The New Vectors model, called BEE (Behavioral Evolution and Extrapolation), is inspired by techniques used to predict the behavior of nonlinear dynamical systems (or chaotic systems). As in the AI simulation, a representation of a nonlinear dynamics system is continually fit to its recent past behavior.
The BEE model is unique as a plan recognition model because it integrates a rational analysis of the agent's goals with both external environmental influences (i.e. sometimes the environment will interfere with the rational plan) and internal emotional biases.
via: New Scientist