An animal in the wild constantly faces choices, such as whether to pursue a meal or flee from a predator. This challenge of selecting an action from many possibilities is fundamental to all organisms. To model this complex decision-making, scientists and engineers use a framework known as a behavior network.
Defining Behavior Networks
A behavior network is a model representing the relationships between different actions an organism or machine can perform. Each node in the network corresponds to a distinct behavior, such as “foraging for food” or “sleeping,” representing the agent’s entire behavioral repertoire. These nodes are interconnected by a web of links.
These connections define how behaviors influence one another and can be either excitatory or inhibitory. An excitatory link means that activating one behavior increases the likelihood of another. For example, the behavior “detecting a threat” would have a strong excitatory link to the “fleeing” behavior, priming the action for execution.
Conversely, inhibitory links represent conflicts between behaviors that cannot be performed simultaneously. An animal cannot sleep and eat at the same time, so the “sleeping” and “eating” nodes would have a strong inhibitory link. When one behavior becomes active, it suppresses the other, ensuring mutually exclusive actions are not attempted together.
The Action Selection Process
The process of choosing what to do next is governed by “activation levels.” Each behavior node has an activation level, a fluctuating value representing the urgency to perform that action. This level rises and falls in response to various inputs, reflecting the agent’s changing priorities.
Activation levels are influenced by two primary sources: internal states and external stimuli. Internal states, such as hunger or fatigue, directly increase the activation of corresponding behaviors. For instance, a prolonged period without food will raise the activation level of the “foraging” node. External stimuli, like the sight of a predator or a potential mate, also contribute to the activation of relevant behavioral nodes.
As these internal and external factors change, multiple behavior nodes can become activated at once, creating competition within the network. Through their inhibitory links, activated nodes vie for dominance by suppressing their rivals. A high activation for “fleeing from a predator” will actively push down the activation level of “drinking water,” even if the agent is thirsty.
This competition leads to a “winner-take-all” outcome. When one behavior node’s activation level crosses a predetermined threshold, that behavior is selected and executed. The “winning” node then sends strong inhibitory signals to all other competitors, ensuring the agent commits to a single action.
Behavior Networks in Animal Behavior
The three-spined stickleback fish offers a clear biological example of a behavior network in action. During the breeding season, a male stickleback must manage several conflicting behavioral drives. He must defend a territory from rivals, court potential mates, and care for his nest.
The appearance of a rival male is an external stimulus that raises the activation of his “attack” node. Simultaneously, the sight of a female increases activation of the “courtship” node. The presence of eggs in his nest provides constant input to the “nest-tending” node, which includes fanning them to keep them oxygenated.
A behavior network model can simulate these choices. Since attacking, courting, and nest-tending are mutually exclusive, their nodes are connected by inhibitory links. If an intruding rival presents a more potent stimulus than a passing female, the “attack” node’s activation will likely reach its threshold first, causing the fish to engage the rival while temporarily suppressing all courtship and parental activities.
Applications in Robotics and Artificial Intelligence
Principles derived from observing animal decision-making are used for engineering autonomous systems. Designers have adopted the behavior network model to build control architectures for robots that must operate in unpredictable environments. This approach allows a robot to arbitrate between its objectives without a complex, top-down script.
A mobile robot, for example, can be programmed with a behavior network to manage its core functions. One node might represent “explore the environment,” another “return to charging station,” and a third “retrieve a designated object.” These nodes receive continuous input from the robot’s sensors, such as its battery monitor, cameras, and navigation systems.
The robot’s decision-making process mirrors the biological model. As its battery depletes, the activation level of the “recharge” node steadily increases. If its camera identifies the target object, the “retrieve” node becomes highly activated. If the battery level drops below a certain point, the “recharge” node’s activation will likely surpass all others, causing the robot to abandon its other tasks and navigate to its charging dock. This architecture creates flexible and life-like behaviors in artificial agents.