Inferring intent from behavior using neural networks trained with an approach-avoidance model

Artificial intelligence (AI) systems use past states of the environment to predict future states of the environment. However, when people make predictions, they also make inferences about the hidden states of other entities, known as Theory of Mind. Accurate inferences about hidden states, such as goals, strategies, and traits, are critical in many situations. For example, a person can identify an aggressive driver changing lanes to make a highway exit and leave extra space in response. AI systems equipped with cognitive models that characterize hidden states from observed behavior may allow for more human-like predictions. To test this, we applied a model of approach-avoidance dynamics to a continuous control task. In this task, participants competed with a computer opponent to achieve five different goals by moving a joystick to control a spaceship in a 2-D environment. On each trial, their goal may require them to collide with the other ship, avoid the other ship, stay close to the other ship, herd the ship to a location, or keep the ship away from a location. Neural networks were trained to predict the goal of each trial from raw behavioral data (i.e., the position of each ship) or from parameters estimated from a model of approach-avoidance gradients. Both networks predicted the participants’ goals well above chance, with overall accuracy outperforming human inference. Further work is necessary to test whether a network trained on both behavioral data and model parameters improves predictions.