1. Introduction

"Here, a "microcosm" means a problem consisting of a wide variety of mutually dependent subproblems, and the subject may not be able to 'survive' (in) the microcosm without letting his major basic functions jointly come into play. ... A microcosm is practically a closed problem insofar as it is a microcosm, and this very closeness renders it a theoretical feasibility. We can obtain the particular set of constraining conditions that makes the subject's behavior optimal in a given microcosm." (Toda 1982).

The concept of micro worlds is closely tied to the concept of autonomous agents:

"It seems certain that, as we understand more about cognition, we will need to explore autonomous systems with limited resources that nevertheless cope successfully with multiple goals, uncertainty about environment, and coordination with other agents. In mammals, these cognitive design problems seem to have been solved, at least in part, by the processes underlying emotions." (Oatley 1987).

Although both concepts are theoretically well elaborated there is little work done that uses micro worlds as complete but simplified environments for autonomous agents. What seems to be missing is the availability of adequate formalisms and tools. This lack of appropriate simulation languages and simulation tools is at least for psychologists an important reason why these concepts are rarely used.
In the next section we give a short overview of the functionality of an implemented software package as an example of what such a tool could look like. In the third section we present an autonomous agent that implements the basic concepts of Toda's Social Fungus Eater.

2. The Autonomous Agent Modeling Environment (AAME)

2.1. Object Oriented Simulation Language

• Regions (e.g., wet zones and other ecological properties)
• Manipulable objects with arbitrary properties (e.g., obstacles, landmarks, food, tools, etc.)
• Attractors (e.g., light sources, odors, dynamic processes like circadian cycles, etc.)
• Agents with different dynamic morphologies, different types of generic sensors, actuators and attractors
• Communication protocols (attractors or blackboards between agents, local and global messages between user and agents)
• Building blocks for autonomous control architectures (structures and processes), e.g., artificial neural networks, cybernetics systems, production systems, ADT programming support, etc.

2.2. Interactive Simulation Instruments

Figure 1. An animated microworld and inspection facilities: 1. weight matrices for ANN; 2. bar displays (here for the range finder); 3. oscilloscopes; 4. perception area of a sensor; 5. network editor.

3. The Social Fungus Eater

Figure 2. Trajectories of a Fungus Eater society

The Fungus-Eater is a fictitious mining robot that is sent to a planet called Taros to collect uranium ore. It uses wild fungi growing on the surface of the planet as the main energy source for its biochemical engine. Little is known about the distribution of uranium ore and fungi. Every activity of the Fungus-Eater, including the brain- computer operations, consumes some specified amount of fungus-storage. If the Fungus- Eater runs out of fungus- storage it dies. Its mission is to collect as much uranium ore as possible, and there is no reward for the amount of collected fungi (adapted from Toda, 1982). Since uranium ore and fungi are usually not found in the same place there is a conflict in the robot's action selection.
We made some further assumptions for the concrete implementation of the Social Fungus-Eaters: They keep a certain distance to each other in order to avoid conflicts at food places and to avoid inefficient mining. On the other hand they also maintain loose contact in this potentially hostile environment, e.g., to help each other in emergency situations. The (emergent) behavior of the Fungus Eater, achieved with our model can be described as follows:
As expected, the agents are mostly found around food sources and mines (see Fig. 2). They gently alternate at food places. Agents with similar hunger conditions build circles of 2 to 5 members. The consumption of fungi is more or less equal for each agent in a circle. A circle breaks up when a certain level of energy is reached or when other agents approach the source. Depending on the parameters, agents leave even rich food regions, e.g. to collect ore or to explore the environment. Often these explorers form a couple (but the partner is irrelevant). Sometimes circles may also be found in places without food or ore. These circles collapse when the consumption motive of one or several members becomes stronger. If they do not find enough food in a reasonable time they die. At the food places, very hungry homecomers push less hungry agents aside and clearly misbehave.

Figure 3. A Model of the Social Fungus Eater

Toda uses cost functions and 'urges' to describe the expected behavior (observer perspective). We modeled this with a cybernetics system (see Fig. 3), using a hedonistic principle based on the energy concept (mechanism in the agent). A Fungus Eater is equipped with remote sensors for ore, fungi and other Fungus Eaters and contact sensors for fungi, ore and obstacles. There is a dynamic set point for the energy level that determines the current hunger condition and thereby triggers eating and approaching food sources. Also dependent on the current hunger condition is the regulation of the social distance to other Fungus Eaters. In situations when adequate sensor signals are missing an explorative behavior is produced. Obstacles temporarily intervene in the ongoing behavior whereas the latter is used as a heuristic for the avoidance strategy.
The control architecture enables the Fungus Eaters to adapt to a dynamic environment. They internalize the current properties of the environment, i.e. the distribution of food and obstacles with their individual energy curves to regulate the ore- food dilemma appropriately.

References

* The preparation of this paper was facilitated by grants of the National Science Foundation (FNRS project 11- 39551.93).