An Architecture for Observational Learning

In this thesis I present an architecture that learns new skills through observation and adapts to the environment through situated experience in the world. Such an architectural growth is bootstrapped from a minimal initial knowledge and the architecture itself is built around the biologically-inspired notion of internal models. The key idea, supported by findings in cognitive neuroscience, is that the same internal models used in overt goal-directed action execution can be covertly re-enacted in simulation to observe and understand the actions of others. The system applies these concepts to learning higher order cognitive functions like learning problem solving skills and social interaction skills. Rather than rea- soning over abstract symbols, the system relies on biologically plausible processes firmly grounded in the actual sensori-motor experience of the agent. The system continuously learns new models and revises existing ones through the observation of other intentional agents in the world and through direct indivi- dual experience. The learning process accumulates knowledge about causal rela- tions between observed events, and acquires complex skills observing and abstract- ing the goals of a demonstrator. To reach its goals, the system exploits the acquired knowledge and uses its internal models to reason about future consequences of its actions. The architec- ture anticipates future needs and perils and through its internal models used in simulation it reasons about the future in a way detached from the current situation. This thesis presents also two case studies used to test the ideas constituting the architecture. The first is a classical AI problem-solving domain: the Sokoban puzzle; the second is the domain of social interaction.