Category learning through action: A study with human and artificial agents

The abilities to learn to categorize and manipulate new objects are closely related and ubiquitous in cognitive agents. To shed light on the underlying mechanisms and the relationship between these capabilities we designed an experimental comparative scenario in which human and artificial agents were asked to learn to manipulate, through a mouse pointer, unfamiliar two-dimensional objects that varied in shape, color, weight (i.e. inertia to the movement) and color intensity. Objects were grouped on the basis of their characteristic and each group was associated with a different target manipulation. The artificial agents learned through a stochastic trial and error process in which variations to the artificial controller were introduced randomly. The controller, an artificial neural network, received as input the perceptual properties of the object and the current position of the mouse pointer and of the barycenter of the object and determined as output the movement of the mouse pointer. The comparison of the behaviors displayed by human and artificial agents allowed us to identify the role of human cognitive biases, mostly related to visual properties (e.g. shape), that may be beneficial or counterproductive for the acquisition of the task. The analysis of the performances instead showed that both humans and artificial agents used overgeneralized behaviors (i.e. they found and performed manipulations that could be applied to objects belonging to more then one category achieving good, although often sub-optimal, performance) and exploited properties that co-determine the effects of agent/environment interaction (i.e. weight) rather than visual properties.