ORGAHEAD is an organizational learning model designed to test different forms of organizations under a common task representation. Each member of the organization receives information from a subordinate or from the environment, makes a decision based on the information and what he or she has learned so far, and provides superiors with an answer to the decision. These decisions are either 'yes/no' or 'yes/maybe/no'. ORGAHEAD falls in the class of intellective models; it is not designed to simulate exact behavior, but capture dynamics, important from a theoretical standpoint.


ORGAHEAD tests a single quasi-hierarchical organization containing up to three levels. These levels are named, from top to bottom, "CEO", "Manager", and "Analyst". The organization does not have to take a pure hierarchical form; however, the information or reporting has to travel upward from the task to CEO. For instance, an Analyst can report directly to the CEO, but can never report to another analyst. Similarly, a CEO can have a direct link to a task bypassing any managers or analysts.

Figure 1. Sample Organization with Labeling for the Levels

Task Activity

The organization performs a classification task. The task information is simply an array or vector of ones and zeros. A predefined decision rule or function performed on the array reveals the true answer, of either yes or no (i.e. 1 or 0) for the task information. A typical decision rule is if there are more 1's than 0's, then the answer is a 1, that is the "majority rules".

Figure 2. Details for Agents, Activities, and Decision Task

A real life classification task often used by Carley as a metaphor is the radar task. An incoming plane is detected on radar and different individuals at the radar station are responsible for detecting different features. Analyst A may detect the speed and direction while Analyst B may be reponsible for whether the pilot has trasmitted a transponder, or identification code. The analysts then report their findings to a superior officer, or using corporate jargon, a manager or several managers. In this real life task, the information can be a number, like speed or direction, or simply yes or no, like did it transmit a transponder code. In ORGAHEAD, the information is reduced to a sequence of ones and zeroes.

Each agent (i.e. analyst, manager, or CEO) receives information from a resource whether that resource is a task or another agent. It compares its information and answers to its superiors based on which answer, of yes or no, allowed the organization to perform well before. The final decision makers, CEOS (yes there may be more than one in this simulation), provide the final and single "organizational" answer. If this answer corresponds to the real answer for the task vector, then all the agents receive positive feedback (i.e. "good job for what you did!"). Otherwise, they receive negative feedback and are less likely to answer the same way when they see the same set of information.

If we want to interpret each bit of the input vector as a separate task, in and of itself, then the mapping from task to the inputs is more subtle. The model does not actually perform the tasks represented by each bit, but performs a decision-making/response task to the input. The model sees either binary or tri-nary instantiations of each input, which could represent the outcome of the specific task or its presence (i.e. its requirement for attention). The organization is responsible for selecting, from a binary set of choices, its response to this situation, set of tasks bits, posed by environment. Each agent responds to a subset of the situation and reports to the its superior its response contingent on its learned belief of the proper response that might ensure organizational success. The criterion for success is a parameter to the model, like the aforementioned majority rule. If more than five instantiations of the task correlates to one situation (e.g. hostility), then the organization must choose one of the binary choices for its behavior in order for its action to be considered a success (e.g. mobilize forces).

The performance, called Efficiency in the program, is simply the percentage of tasks for which it provided the correct answer.

Life Cycle

The organization is simulated over a pre-determined number of tasks generally several thousand. "Absolute Efficiency" (i.e. overall performance) is measured as well as the recent efficiency, called "Relative Efficiency".

Organizational Adaptation

Every so often, a change is randomly proposed to the organization (see Figure 2, Organizational Activities). This change can be a hiring of a new agent, firing, or changing linkages, either from agent to task or agent to agent. The organization is capable of estimating its short-range performance as a result of accepting this change; it can "look-ahead", which is where the "AHEAD" part of ORGAHEAD comes from. It will always accept a change that is estimated to improve performance and sometimes accept those that it thinks will temporarily degrade performance. The mechanism that determines whether it will accept bad, or "risky", changes is the optimization heuristic, simulated annealing.

Simulated Annealing

Simulated annealing is an optimization heuristic. By optimization heursistic, we mean an algorirthm designed to solve a really difficult problem for which the best answers are computationally difficult to assess. Optimization heuristics instead provide very good answers, not always the best. You may have heard of other such algorithms, like the genetic algorithm (GA) or hill-climbing or A-star.

Metropolis Criterion and Simulated Annealing

Simulated annealing is a heuristic borne of metallurgical annealing. Annealing is a process in which alloys are slowly cooled to obtain an optimal molecular configuration. Cooling the alloys too fast results in a less stable configuration. In the computer world, this process is interpreted as the process of solving the problem by allowing for moves, or step, which doesn't necessarily seem to take us closer to the solution. The reason for this is that, in many complex problems, you need to take a few bad steps in order to find the path that leads you to the best or "very good" answer.

Think about a blind man trying to reach the highest point in range of hills and mountains. His cane will only tell him when he's reached a peak, but whether that peak is the highest of them all cannot be assessed. In ORGAHEAD, the simulated annealing process is meant to be a metophor for an organization's "liability of newness". That is, a new organization needs to be open to taking some risks, or else it will never learn the appropriate strategy. Over time, the organization will have learned enough to not need to take the risks. At the heart of the annealing process is the Metropolis criterion, or the probability to accept a bad move constrained by a "temperature" variable:

probabilityt = e-costt*k/Tempt (Metropolis criterion) [1]

Tempt = a * Tempt-1 where 0.0 < a < 1.0 (temperature cooling) [2]

costt = current_performancet - lookahead_performancet (example cost calculation) [3]

Figure 3. Cooling shown as a Probability of Accepting Costly Moves, measured empirically

Virtual Experiments

Virtual experiments are run to address a variety of questions regarding organizational learning and performance including the following:

Sample of Results

Some of the analyses have focused on understanding the differences between organizations that evolve to perform well and those that do not; we call these adaptive and maladaptive organizations, respectively. The following figure depicts the activity strategies for a set of ten adaptive organizations compared to ten maladaptive ones. Each letter represents an accepted change at each change cycle: H for a single hire, F for a single fire, and T for a structural change (i.e. tie change).

Figure 4. Activity Sequences for Adaptive and Maladaptive Organizations.

One of the prominent differences in activities is the rate of firing, being higher for maladaptive organizations. Not surprisingly, maladaptive organizations tend to be smaller than adaptive organizations. More subtle differences in activity can be determined by examining the subsections of the sequences. For instance, which combinations of hire, fire, and tie change produces marginal gains in performance? Is there a strategy involving firing that can produce higher performance (e.g. by firing the ineffective individuals)?

The following link pops-up, in a single window, a set of animated plots which depict the activity cycles of adaptive and malaptive organizations. Each axes represents the mean probabilities of the orgnization performing a hire, fire, or tie change and the plots depict each combination of activities (hire vs. fire, hire vs. tie change, and fire vs. tie change). The flow of particles represents the change in the activity over the life cycle of the organization. The width of the flow represents the standard error of the means.

NOTE: The animation loops. Press <Esc> to stop animation. Press Ctrl+R to restart it.

Click Here to Compare Activities of Adaptive and Maladaptive Orgs

The graphs are quite telling. The lower right graph shows a back and forth hiring/tie change cycle for adaptive organizations. Maladaptive organizations start out with a greater probability for firing and, while that decreases, it's hiring and firing activities don't achieve the same rates as the adaptive organizations.

Support is provided in part by the National Science Foundation IGERT in CASOS, the Army Research Labs - study on "Personnel Turnover and Team Performance", and the Office of Naval Research ONR 1681-1-1001944.

Views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Labs, Office of Naval Research, the National Science Foundation, or the U.S. government.

Web Page Author: Ju-Sung Lee,
Last Update: 9/7/01