Introduction
In the late 1980s, Artificial Intelligence faced a significant challenge: how to enable autonomous agents to make decisions in complex, dynamic environments. Traditional centralized planning methods were often too slow and brittle to handle real-world unpredictability. In her seminal paper, Pattie Maes proposes a decentralized alternative: a system where rational behavior emerges from the dynamics of competing "competence modules." This approach allows agents to remain goal-oriented while being highly adaptive and robust.
Problem Statement: The Action Selection Dilemma
The "action selection" problem involves deciding which action to perform at any given moment to satisfy multiple, often conflicting, goals with limited resources. For instance, a robot might need to sand a board and paint itself using only two hands. Classical AI typically managed these goals through a central controller and rigid search trees. However, such systems often fail when the environment changes unexpectedly—for example, if a board slips out of a robot's grasp. Maes argues for a distributed system that is opportunistic, robust, and avoids the bottlenecks of centralized control.
Methodology: Spreading Activation Networks
Maes conceptualizes an agent as a "society" of interacting competence modules, inspired by Marvin Minsky’s Society of Mind. Each module is defined by its preconditions (the state required for the action) and its effects (the resulting state change).
The core of the system is a Spreading Activation mechanism:
- External Input: Modules receive activation from the current environmental state and global goals.
- Internal Spreading: Modules are linked in a network. Executable modules (those with met preconditions) spread activation forward to successors. Non-executable modules spread activation backward to predecessors, signaling a need for their preconditions to be satisfied.
- Thresholding: At each timestep, the module with the highest activation level exceeding a specific threshold $\theta$ is selected for execution.
Experimental Evaluation
The proposed model was implemented in Common-LISP and tested across three distinct simulation environments:
- Minsky’s Tower: A 7-agent society tasked with block-stacking.
- The Two-Handed Robot: A 10-agent society coordinating two hands to sand a board and paint itself, requiring the use of a vise.
- Conflicting Blocks World: A complex 18-agent network with 594 links designed to test interacting goals.
Results and Key Insights
The simulations demonstrated that rational behavior could emerge without explicit programming. The "plan" is not a static data structure but a dynamic pattern of energy within the network. This allows for "partial replanning" at every timestep, making the system exceptionally resilient.
Key findings include:
- Behavioral Tuning: The agent's "personality" can be adjusted via global parameters. Increasing the activation threshold makes the agent more deliberative, while increasing environmental influence makes it more opportunistic.
- Performance Trade-offs: Maes identified qualitative laws governing the system. For example, higher thresholds generally lead to more optimal action sequences but decrease the overall speed of decision-making.
Limitations and Future Directions
Despite its strengths, the model has limitations. The agent's look-ahead is limited by the reach of activation spreading, and the lack of memory means it may repeat planning errors. Additionally, about 8% of simulation runs resulted in "impasses" where no agent reached the firing threshold. Maes suggests that a hierarchical "second society" of agents could monitor and tune these parameters in real-time to mitigate such issues.
Summary updated