Simulated Ant Colonies

Simulated ant colonies form a helpful example for multi-agent simulations, as their communication works based on indirect interaction. In this paper I describe a framework that was developed for the simulation of ant colonies, but can also be applied to other domains. Starting from the overall structure in the multi-agent model, followed by a description of primitive framework components, activities and roles are introduced as means for structuring behavior description. The result is a framework for explicitly representing ant-like multi-agent models that scales also for more detailed model. A model represented using this framework can be interpreted for simulation. A complete environment for modeling and experimenting was build around it. The application domain was mostly ant simulation, but also for models in other examples, the framework was successfully applied. Evolutionary models and Multi-Agent Simulation Analyzing emergent phenomena and patterns requires the modeling and simulation of multi agent systems. Emergent phenomena are characterized by locally interacting entities that produce a pattern or behavior observable on a global scale that is not directly deducible from local behavior [1]; shortly stated: “the whole is more than the sum of its parts”. Ant colonies are a good example for such systems. Although the local behavior of a single ant does not seem to be controlled centrally, nor any explicit coordination between ants is observable, the superorganism “ant colony” is able to construct complex nest architectures or adapt its distribution of foragers to food sources in an efficient way. Emergent phenomena can in general be found in complex adaptive systems, where every entity adapts to its local perceptions and by its interaction triggers other entities also to adapt their behavior [2]. Evolutionary multi-agent models can reproduce complex adaptive systems. A multi-agent model is a simulated multi-agent system that exists in a simulated environment [3]. It is a natural form of modeling, especially for societies, as active entities in the original are also interpreted as active in the model. That means that systems, which can be observed as multi agent systems in the original, are also conceived as multi agent systems in the model. Evolutionary multi-agent models additionally capture some process of development in the agent system, not necessarily but mostly with the help of genetic operators, thus reproducing natural evolution. However, also simulation of individually learning agents may belong to that class of evolutionary models. The modeler has to control both, multi-agent system – behavior and development – and environment. The latter is especially important in evolutionary models as it determines the direction of the adaptation. Multiagent simulation is an appropriate method, when in-homogeneities in space or in the individuals are important. Thus it is especially useful for social simulation incorporating mental models into the single actors. But these models are not only costly, but also difficult to design, due to the normally complex form of interactions. However, the example of ant colonies shows up to be sufficient and practicable for many application areas, although every agent has very simple structure. Simulated Ant Colonies as an Example Simulated ant colonies are a good example for multi-agent models. Their supposed form of communication is indirect interaction, or “stigmercy” [4]. That means, that one agent that perceives the configuration of its surroundings and its internal state, changes its environment e.g. by depositing some pheromone trail. Other agents perceive the same environment including the trail and react on it, e.g. by following it. Thus in an ant model there is no need to incorporate complex direct interaction protocols. Stigmercy interaction is simple, but has important properties: There are no hard-wired channels of interaction. Thus the agent behavior must be flexible and adaptive. However, it can be formulated rather independently from the behavior of the other agents. It is just depending on the perceived environmental configuration. The main drawback is that the agent behavior must be synchronized indirectly. The local behavior that should result in the global phenomenon under examination is modeled without formulating direct interactions, but based on some sequence of environmental manipulations. The possibility to focus on single agent behavior is useful for formulating evolutionary models: as the individuals with such behavior models are due to selection, interaction protocol with direct messaging would be more sophisticated to deal with. Frameworks for Multi-Agent Models Even without an explicitly incorporating mechanism for evolutionary adaptation, the design and implementation of a multi-agent simulation is not trivial. The concept model is easily accessible, as active components in the system – i.e. the agents – are also active components in the model. But having to deal with concurrency or synchronized behavior is not as easy as it seems at first sight. Thus support by appropriate development environments is necessary. We developed a framework especially for ant-colony-like multi-agent models. Our goal was to support domain experts to an extent, that they are able to formalize their models themselves. There are already lots of frameworks, tools, and languages for multi agent simulation. Approaches range from schemata and systems originating in the simulation area to languages for modeling multi-agent systems developed in Distributed Artificial Intelligence. Multi-agent simulation can be seen as an enhancement of standard individual oriented simulation. Agents can be taken as equipped with more sophisticated behavior and enabled with methods of adaptation, in comparison to the individuals in process-oriented models or cellular automatons [5]. In analogy to viewing agent-oriented systems as a more abstract form of object-oriented systems, multi-agent simulation can be related to object-oriented simulation [6]. In the last years a variety of multi-agent simulation languages, e.g. Swarm [7], agent architectures for simulated agents like PECS (“Physis, Emotion, Cognition, Social Status”) [8] or specification frameworks like AgedDEVS (“agent-oriented DEVS”) [9] were developed. However, all of them require direct programming, as they do not provide an explicit model representation. On the other hand, there are declarative languages and schemata for describing multi-agent system based on parallel rule interpreters, like SDML (“Strictly Declarative Modelling Language”) [10] or special forms of logical representation, like ConcurrentMetateM [11]. These frameworks provide a completely declarative, rule-based modeling framework, but support for simulation application does almost not exist. Rule-based representations are especially apt for describing behavior. However, these frameworks possess almost no means for structuring the rule sets. Thus one can suppose that modeling more complex models is a rather hard task. Summing up, we can identify two major drawbacks in existing frameworks for ant-colony-like multi-agent models: Totally explicit modeling languages have a scalability problem. Although rules provide procedural semantics, but declarative representation [12], grounding of truth-values of checks and action effects is problematic. This is due to the fact that a complete declarative representation leads to large rule sets, as also all basic principles have to be formalized. On the other hand, a huge amount of class libraries, simulation languages base on non-declarative model representation. The model is coded in a programming language without separating the model from simulator or experimental control. Due to the a priori higher level of abstraction that becomes possible using an explicit, i.e. declarative form of representation, this separation has many advantages [13, 3]. One of them is the possibility to provide additional specialized tools for a concrete representational scheme. Examples are appropriate graphical user interfaces for modeling and experimenting. The second drawback is directly connected to that large rule sets. In general, large unstructured rule sets have proven as impracticable for large knowledge-based systems [14]. As we want to develop more detailed models for individual agents, the resulting disadvantages can be transferred to the description of agent behavior. Lack of clarity in un-structured rule sets: Editing a large rule set becomes rather expensive not only because navigation takes a lot of time. Similar rules, conflicting rules, etc have to be found and eliminated or modified. Costly formulating of rule conditions or conflict resolution? The consequence is, that either one accepts the effort maintaining a rule base or transfers the problem of similar or conflicting rule to the conflict resolution. Then the interpreter selects all rules which’s condition is true and the conflict resolution hast to cope with more than one rule firing. Difficult generation of coherent behavior: This is a consequence from the first point. If there are many rules with similar condition, but different action, the situation changes slightly, but a different action is selected. When the modeler looses his overview over the rule set, the generate behavior become