ml-lab / dual-coding-agents

A Dual-Coding Architecture for Intelligent Social Agents

A Dual-Coding Architecture for Intelligent Social Agents

---

1. Overview

1.1) Agent System

This paper describes a cognitive architecture for intelligent agents who communicate with one another in a social environment, while at the same time performing actions in a situated or physical environment. Agents therefore perceive and act within two domains: physical and social.

1.2) Situated Cognition

Physical behavior is carried out by a system of motors, and social behavior is carried out by collection of signals. Motor outputs manipulate the agent’s spatial position and signals alter its position in society. Certain physical and mental properties acting on the agent ultimately determine the effect of any given action. A real world example of this can be seen in the way we are able to move ourselves forward in space by performing certain leg-muscle patterns which evoke frictional forces on our feet. The transfer of energy caused by the attempt to push back and downward on the ground accelerates the upper body in a forward direction.

Friction, of course, does not depend on the person walking, nor does the person have any control over the nature of frictional forces. The act of walking is simply the result of applying knowledge regarding certain actions in various contexts and how they change the environment, in order to satisfy a set of goals actively being pursued by the person, i.e. to change their position in space from the current position A to a desired position B. Similarly, agents must learn the specifics of their environment in order to succeed in achieving their goals.

1.3) Social Importance

Social behavior is carried out by a system of signals received by other agents. Certain signal patterns carry with them a symbolic relevance to an individual or a collective population of agents. A real world example of this is seen on websites like twitter, where memes develop highly significant roles in the population, before it inevitably dies out and is replaced by another, in a seemingly endless loop. Obviously, a meme, like any idea, does not require every member of the population to have direct access to the original source, because it is spreads between nodes of the social network and encodes itself in the memory of each member’s brain.

Similarly, agents transmit ideas back and forth in order to learn from others as well as teach them. The certain mental processes that result in this communication mechanism are beyond any agent’s control, but one may learn how these mechanisms work in order to increase its control over itself and the environment.

 

2. Perception

2.1) Sensory Reception

Agents receive inputs from a two-dimensional array of receptors from their physical environment. Each receptor holds a value between 1 and 0, which depends on the environmental input to the system. The receptive field has structure-preserving connections to a set of nodes, which are specialized to detect contrast over space.

Nodes can take form as one of two individual types of contract detection, called on-center and off-center. Each type responds to relational patterns between center and surrounding receptors in a unique way. The response patterns of each type are shown in the table below.

2.2) Symbolic Interaction

Agents receive signals one a one dimensional array of receptors from the social environment. Each receptor holds a value between 1 and 0, which depends on the frequency of the input and the associated range of the receptor.

The receptive field has structure-preserving connections to a set of nodes, which are specialized to detect contrast over time. Each node is assigned a frequency range and will detect if the input signal falls within that range.