Second-order networks for wall-building agents

This paper describes robust neurocontrollers for groups of agents that perform construction tasks. They enable agents to balance multiple goals, perform sequences of actions and survive while building walls, corridors, intersections , and briar patches. 1 Purpose We are developing neurally-controlled agents that can build structures out of building materials found in a two-dimensional continuous environment consisting of circular objects called discs. The agents must complete the sequential building tasks while performing other tasks necessary to their survival, such as eating, drinking or eeing predators. In order to perform these construction tasks properly, it is important that our agents' neural control mechanisms exhibit: balance between multiple goals, robust sequencing, backtracking, connict resolution and opportunism 6]. An agent must be able to perform long sequences of actions. For instance, when scavenging, an agent must look for and nd an object to scavenge, go to it, grasp it, look for and nd the location to take it to, go there and drop the object. An agent needs to detect problems in a sequence, including when the agent fails to complete some action or when some other prior state of a sequence becomes undone. When these problems are detected, the agent needs to return to an earlier step in the sequence to correct the problem. For example, if a scavenging agent is carrying a disc back to the scavenge-site when it stops to drink, it must drop the disc it is carrying in order to drink. The agent has to know that it's no longer carrying a disc and must go nd another one before returning to the piling site. An agent should be able to make adjustments in behavior in order to take advantage of opportunities. For instance, if an agent is carrying a disc to a destination and happens to pass near food, it should eat the food encountered along the way. This paper describes a heterarchical architecture of second-order neural networks that enable agents to perform construction tasks while exhibiting the above properties. 2 Task Algorithms In order to test the agents' controllers, the agents need speciic construction tasks. For these tasks, the agents need algorithms that are executable by the controllers. This section will describe the tasks and the algorithms the agents use to complete them. The agents had three construction tasks to complete: building straight walls; building corridors and intersections; and building briar patches. The agents are handicapped …