Research Issues for Biological Architectures in Sensor Networks

1. Coordination of activity among multi-functional agents

Since the sensor network may consist of agents with different functional abilities, the coordination of activity in space and time is an open research question. Should all agents have explicit knowledge of the existence and capabilities of other types of agents?

For example, mapper agents may explore the territory first while the mobile lab agents remain asleep. When the terrirtory has been explored sufficiently, the mapper agents will wake up the mobile lab agents and direct or lead them to the most promising areas.

If agents know the capabilities of other agents, it will result in a more integrated system. However, the system will also be more complex and less flexible to changing missions. Thus, possbile criteria to use when deciding between the two approaches are ability to achieve the goal, complexity, and flexibility.

2. Spreading or sharing of knowledge gained by the agents

Agents may need to spread or share knowledge that they aquired. Should such knowledge be globally distributed, globally distributed with various levels of detail, partially distributed, or not distributed at all?

As an example of globally distributed information, mapper agents may explore an environment and transmit all data to all other agents. Thus each agent has a complete map of the territory. As an example of not distributing the information at all, each mapper may drop simple pheromone markers at various locations in the environment. Thus, no agent has a complete map of the environment.

Some data samples may be wrong because of faulty agents. Thus data samples may need to be weighted (i.e. if 3 agents report the same reading, then it is more reliable than if only 1 agent reported the reading.)

Again, possible criteria to use when deciding among these approaches are efficiency of achieving the goal, memory requirements, and complexity.

3. What is needed in the communication infrastructure

Pheromones are commonly used in biological architectures for communication purposes. Obviously, it is very difficult to faithfully duplicate pheromones in machine architectures. However, it is possible to list characteristics of pheromones and then chose those that are needed to accomplish a goal.

  • presence or absence of pheromone: This allows agents mark a location, or to lay a trail, similar to ants.
  • propagation of the pheromone through the environment: This allows agents broadcast simple or rich information to a group of nearby agents.
  • decay of the pheromone through distance: This limits the distance that information is broadcast to reduce information overload on the agents. It also allows the location of an agent by following a trail of increasing strength.
  • decay of the pheromone through time: This acts as a cleanser of stale or sub-optimal state from the system. Decay of pheromones through distance and time makes the strength of the pheromone an additional souce of information in addition to the presence or absence of the pheromone itself.
  • aditional bits of information encoded in the pheromone: This allows agents to convey additional information, such as danger, type of agent, status, etc.
  • secure pheromone generation and propagation: This is needed in environments where security is a concern. Only valid agents should be able to generate a valid pheromone, not decoy agents or agents that have been tampered with.
  • immediate detection of a pheromone: If a pheromone is immediatly detectable, then an agent can make its decision rapidly. If a pheromone is only detectable once every 30 seconds or longer, then an agent's decision time is similarly increased.

    The implementation of pheromones strongly depend on which characteristics are needed and the physical environment. If only the presence/absence of pheromones are needed in a physical environment, then metal pellets or shavings from a magnetic material may be sufficient. More characteristics can be supported using low-power radio frequency transmitters. Because software is extremely flexible, all of the above characteristics can be implemented in an operating system (AgentOS).

    4. Reverting to a safe state and dynamic software reconfiguration

    Due to catastrophic unforseen circumstances, agents may need to revert to a safe state and await new instructions from a distant controller. The definition and implementation of a safe state is an open research question. A safe state that was defined when the system was initially built may no longer apply. The agent may need to continually define or establish new safe states. Safe state may be a system property as well as a property of an individual agent. Even if all agents revert to an individual safe state, the system may not be in a safe state.

    (Is there an example of this?)