Software agents that learn through observation

Abstract

In this paper, we present an architecture for software agents that enables them to learn vocabulary through the observation of each other bodies and actions. Besides sensors, effectors, and action control, the architecture provides the equivalent of a body with visual appearance. The agent soft visual appearance is designed to be seen by other software agents, not by people. The paper describes the agent software body with visual appearance, the learning mechanism, the demonstration scenario and presents some results showing that the agent software body allows agents to learn vocabulary through observation and to ground the meaning of the symbols they learn. The paper emphasizes the important role of social interaction in learning processes.

Learning by Observation of Agent Software Images

“…As section 2 shows, to the best of our knowledge, the only approach to learning by observation in software agents is the one presented by Machado and Botelho (2006). However, in Machado and Botelho's proposal the agents were only capable of learning vocabulary whereas in the work described in this paper the agents are capable of learning control mechanisms, which requires developing a different kind of learning algorithm.…”

“…The literature overview regarding learning by observation shows that, with the exception of Machado and Botelho's (2006) approach, software agents are disregarded from the advances on learning by observation since they are usually related to robotics (Argall et al, 2009). Software agents are not able to distinguish themselves or other software agents from the remaining elements of a computer program because of the disembodied nature of software.…”

“…This section presents a summary of the additional functionalities that our approach introduced to previous work regarding the software image (Machado & Botelho, 2006). The section describes these new functionalities, explains the reasons for their inclusion in the software image and how they are advantageous for learning by observation.…”

“…Figure 1 shows a representation of the key elements of the software image. Like in Machado and Botelho's (2006) proposal, the elements of the agent software image are arranged in two categories, the static image and the dynamic image. The static image is immutable (does not change over time) and describes the constituents of the agent whereas the dynamic image changes with time and describes the activities of the agent.…”

“…As with learning by observation in humans and superior mammals, it is the apprentice who takes action to obtain new knowledge, and does it without interfering with the expert agent. In spite of the similarities with Machado and Botelho's (2006) approach, our approach introduced several improvements to the software image, in particular:…”