The overall objective of the MATHESIS project is to explore fundamental aspects of social communication and adaptive behaviour, namely the process of assigning meaning to the actions of other subjects. This will be demonstrated by developing artificial cognitive agents able to acquire a motor repertory by observational learning. Observational learning is understood here as the capacity to acquire an action strategy through observation only, without the experimentation needed in learning procedures, which mostly use trial-and-error (e.g., reinforcement learning). Observational learning is also distinct from imitation, a term that we will use to refer to agents enacting a strategy soon after observing its use by another agent. This overall objective can be divided into four primary objectives (P1-P4):

P1. Elucidate the mechanisms responsible for observational learning in biological agents (humans and non-human primates) by collecting relevant experimental data (from neurophysiology, neuroanatomy, functional brain imaging and psychophysics). Perform experiments to reveal the cortical network involved in: (i) The execution of actions (visually and memory guided) and (ii) the observation of the same actions performed by another subject. Recent related results indicate that executed, observed and memorized actions are represented in extensively overlapping parieto-frontal cortical networks of the primate brain. We intend to demonstrate conclusively that the same distributed cortical network is activated both when subjects act, as well as when they observe or recognize the same actions executed by others. Moreover, we intend to explore the developmental path which endows healthy infants and autistic children with the ability to learn from observing another person perform an act, formulate a cognitive model of observational learning, and draw conclusions about the efficacy of mental training in children. Focusing on autistic children is of particular interest, in that it will allow us to compare both the similarities and the differences between observational learning and imitation, and to design appropriate educational interventions. Observational learning is claimed to be specifically impaired in children with autism, whereas imitation is normal, as long as familiar actions are concerned. Thus, the communicative role of imitation can be activated in such a way as to contribute to the build up of a motor repertory of actions enabling the children to learn new actions via observation.

P2. Construct a general computational model for observational learning, guided by the biological mechanisms revealed in P1. Extract the pertinent information-processing principles and formulate a computational model suitable for implementation on artificial agents. The computational model will be based on a biologically inspired block diagram outlining presently available evidence about the connectivity of the cortical areas involved in observational learning, and the task-specific information they convey to each other.

P3. Produce embodiments of simple cognitive agents (with basic perceptual, motor and mnemonic abilities), equipped with the computational model of observational learning developed in the context of P2. The embodiments will have the form of robotic and software agents, and will be used to evaluate the developed computational model. The robotic agents will reproduce the developmental path, unveiled in P1, which endows infants with observational learning capabilities, and will be used to explore mental training in robotics. The prototypical motor actions considered for the robotic agents will be reaching for objects, as well as grasping, under the guidance of appropriate visual, haptic and proprioceptive information. The robotic agents will be used to assess the performance of the observational learning model developed in P2, in terms of robustness, accuracy, generality and efficiency. A second embodiment of such a cognitive agent will take the form of a software agent in a suitable application domain (e.g. competing social groups of artificial "beings" populating Artificial Life worlds). Both types of embodied cognitive agents will be used to develop a detailed evaluation framework for the computational models of observational learning and mental training.

P4. The embodied cognitive agents, able to learn by observation of other agents’ actions, will be based on models and methods from neuroscience and psychology. Conversely, these agents will be used to refine the questions posed by neuroscience and psychology regarding the mental training capabilities of biological agents. The recent neuroscientific results mentioned in P1, for example, bring up issues which can be difficult to address by anatomical or imaging methods alone. The embodied cognitive agents of P3 will provide a means of further exploring some instances of these issues, thus becoming important tools to the study of biological agents.