Policy Evolution within an Organization

A plan of action is given for a newly funded research project on organizational evolution. In this study, our goals are to apply an evolutionary framework to organizational learning. The process includes a collaborative effort with partner companies to identify the working mechanisms behind the evolution of policies or decision rules. We also use computer simulations as a tool to examine our findings. Introduction: This document describes a plan of action for a three-phase research project funded under the Innovation and Organizational Change program by the National Science Foundation in organizational learning using an evolutionary framework. The long-term objective of our ongoing research is to explicate principles that govern evolution within an organization. On a more practical side, we hope to provide managers with a set of rules or guidelines that will permit their companies to evolve more rapidly in desirable directions. These rules might deal with the appropriate number of business units, promotion policies, or team-based decision making, among other organizational characteristics [1]. Our first-phase goal is to perform evolutionary audits– that is, to collect data on the evolutionary potential of several partner companies. The evolutionary audits will aid in the development of a model of organizational evolution (phase II). The data we collect and the model that we build will form a foundation for in-depth explorations of the question: How can managers create an organization that will evolve quickly in the direction they desire? (Phase III) [2]. Partner companies will be involved in all three phases. Our partners include PriceWaterhouseCoopers, Pugh Roberts Associates, Eastman Chemicals, Hewlett Packard, The Lincoln School and General Motors. We selected these particular companies for their interest in this work, their existing expertise in one of the underlying methodologies (system dynamics), and certain organizational characteristics (discussed below) which make them particularly good subjects for studying organizational evolution. Justification: Two observations suggest this effort is worthwhile: First, we are addressing a significant problem, and, second, the problem is solvable. The problem is that system-wide company improvement is difficult because companies are too complex to “solve” [3]. Creating a consistent set of beneficial management policies is difficult or impossible because the complexity of modern companies exceeds, by many orders of magnitude, our ability to understand. Managers today work on isolated issues that they can identify. Sometimes the issues are solved, but often this reactionary approach leads to unintended consequences, as intended solutions create problems in other areas of the organization [4]. The problem of improving organizations in the face of ignorance is solvable. In fact, it has been solved, just not by humans. Biological evolution has produced excellent natural organizations (i.e. organisms) even though the organizations themselves are completely ignorant of how they are put together or why they succeed. We are identifying analogs of natural evolution that will help companies to likewise excel. In this paper, the foundations of the study, the technical methodology, and our current status are presented. Our research plan follows this background material. A schedule follows the research plan. Foundation: Central to our work is a particular analogy between biological and organizational evolution. Analogies can be dangerous when careless application leads to unwarranted transfer of conclusions from one domain to another. Properly employed, however, analogies are powerful mechanisms for using precious knowledge from one area to bootstrap understanding in another [5]. Today, evolution is increasingly seen as a general mechanism, not restricted to the biological realm [6-8]. Still, the evolutionary 1 The difficulty of making organization-wide improvements clearly increases the value of an evolutionary approach. However, this does not mean that evolutionary management is mutually exclusive of other improvement efforts. Indeed, good evolutionary mechanisms will tend to spread any beneficial change whether the change is intentional or not. The same evolutionary mechanisms will limit the spread of deleterious changes, again whether or not the changes are intentional. 2 Interestingly, Darwin himself was lead to his theory of evolution partly through an analogy between Adam Smith’s political economy and biology. And, Adam Smith had been influenced by an analogy between Newtonian physics and political economy [9]. mechanisms we understand best are biological. Hence, in seeking to understand how an organization can evolve it is natural to seek fruitful and powerful analogies between the biological and the organizational. A focal point of our research has been an analogy between organizational policies and biological genes. By policy we mean an explicit or implicit decision rule in the usual system dynamics sense [10]. For example, a manager might set prices by the implicit rule: Raise prices when inventories are low, and lower prices when inventories are high. The policy is implicit as long as it remains unspoken or unwritten. Articulating the policy, perhaps by recording it in a policy manual, could make it more explicit. Of course, people might change their approach to pricing even without updating the manual. In this case, the new approach would be a policy, while the old procedure recorded in the manual would no longer be a policy in our use of the term. A policy is a rule or procedure that people actually use to make a series of decisions. In this case, the policy gives rise to a continuing stream of particular decisions to raise or lower price. A policy in an organization is comparable to a gene in a cell. A gene is a segment of DNA (or, in some organisms, RNA) that acts as a set of instructions for the ongoing production of a particular protein. The proteins then catalyze reactions in the cell. Indeed, no necessary chemical reaction occurs in a cell without a protein catalyst that is coded by a gene [11]. Genes produce a continuing stream of action in the cell, while policies produce a continuing stream of action in the company. The creative mechanisms in evolution are mutation and recombination. In genetic mutation, part of a DNA molecule is physically changed, producing a new gene. In our analogy, genetic mutation corresponds to policy change, intentional or unintentional, and (like mutation) producing either favorable or unfavorable results [12]. The result of such a change, for better or worse, is a new policy. Genetic recombination occurs when two DNA molecules mix to form a new DNA molecule. In a company, genetic recombination corresponds to a particular kind of organizational learning: Inter-personal learning whereby a person combines a part of someone else’s ideas with his or her own [5, 13]. Evolutionary management consists primarily of managing the environment in which policy change and learning (mutation and recombination) occur. One important task is to create mechanisms that will ensure the spread of more effective policies and the decay of less effective ones. That is, managers need to create mechanisms that correspond to nature’s processes for the spread and selection of genes. In higher animals, sexual reproduction encourages the dissemination and recombination of genetic material. Natural selection is the process by which beneficial recombinants (and mutations) are retained, while deleterious ones are discarded. Sex and selection in the natural world correspond in the corporate world to the various ways in which companies identify certain employees as exemplary and encourage other employees to learn from (or imitate) them [14-19]. We call these processes of identification and encouragement “pointing and pushing mechanisms”. For example, pay and organizational position are two ways that a company can point to outstanding performers. Pay and position can serve a pushing function as well: Employees are motivated (“pushed”) to learn via their desire to rise in the pay scale and in the hierarchy. Other pointing and pushing mechanisms are also possible. 3 Pushing successful people to share policies with others may also be necessary [20]. Methodology: We have developed theory and preliminary formal models of organizational evolution to deepen our understanding and to gain additional insight. Our approach to modeling has been to combine system dynamics modeling [10, 21], agent based modeling [22, 23] and genetic algorithms [2426]. System dynamics models are built around nonlinear feedback processes. The model formulations are similar to those found in feedback and control models of electronic, mechanical, and biological systems, but are used in system dynamics usually (though not exclusively) to investigate human-related systems. The system dynamics methodology has been applied to a broad diversity of topics in organizational dynamics including project management, inventory supply chain management, environmental systems management, and urban planning [10, 27-31]. Here, we use a system dynamics model to simulate the progress of projects within a company, given a set of policies. Agent based models are built around simulated agents –or, in our case, managers. Each agent has a repertoire of behavior and can interact with other agents. The key behavior of our agents (managers) is making decisions based on policies and the key interaction is learning from another manager and thereby changing the policy. Genetic algorithms are computer programs that solve problems by mimicking the biological process of evolution [24, 25]. A standard genetic algorithm would begin with a population of potential solutions to a problem and would evolve ever-fitter solutions through processes of mutation, recombination, and selection. Genetic algorithms have been applied to a wide range of optimization problems including microwave antenna design, circuit design and airline scheduling [32-35]. In most applications of this technique, the genetic algorithm is simply used as an optimization technique and is not intended to represent a real-world process operating within the problem domain. In our work, however, we interpret the genetic algorithm as representing a particular type of human learning – the process by which one person incorporates ideas from another into his or her own policies. For example, perhaps a manager (agent) follows the previously mentioned policy of raising prices when inventories are low and lowering prices when inventories are high. In this way his price effectively responds to excess supply (high inventories) or demand (low inventories). Perhaps he comes into contact with another manager with a different policy, a manager who sets prices by taking costs and adding a fixed margin of two hundred percent. The second manager’s policy ensures that prices never fall below costs. From this contact, our original manager might learn the idea of margin pricing. Henceforth he might price at a margin over cost, like the second manager, but in addition, he might vary the margin depending on inventory position, similar to his original policy. The new policy is actually a combination of two parent policies, combined by learning. In this example, the new policy might perform better than either original policy, because it responds to inventory position (supply) and at the same time ensures that price will never fall below cost. Generally, we use system dynamics, agentbased modeling and genetic algorithms in the following way: A system dynamics model represents the underlying physics of the organization as well as all policies that are not 4 Of course the recombination can be detrimental as well [36]. evolving. Evolving policies are carried by agents (individual managers) who learn from one another via a process that is essentially a genetic algorithm. The completed simulation environment will allow us to investigate the conditions and trade-offs that influence the rate at which organizations improve their policies. Conditions include number of teams, team size, frequency of mixing and evaluating, promotion (or other pointing and pushing) policies, number of managers, complexity of evaluation criteria, complexity of policies, number of policies, and many others. An example of a trade-off is between the ability to discriminate between good and bad policies on the one hand and, on the other hand, the speed by which a policy spreads. As the number of teams increase, the rate by which policy changes disseminate will slow, while the ability to discriminate between beneficial and deleterious changes will increase. One important contribution of our methodological approach will be a tested method for building simulation models that will enable managers to investigate the efficiency of proposed or existing evolutionary mechanisms in their own companies. Current status: We have initiated this study by developing a preliminary theory of organizational evolution and creating a proofof-concept model to explore and deepen our theory of organizational evolution as well as to demonstrate the feasibility of combining system dynamics with agent based and genetic modeling. Throughout, we have sought examples from the business sector to support, clarify, or debunk our findings. The proof-of-concept computer simulation environment coordinates two modules: a system dynamics model and an agent-based module operating under a genetic algorithm. More specifically, we have created a simple system dynamics model of a company that simultaneously runs a number of projects (shown in Figure 1). We interpret the equations in terms of a software company or automobile manufacturer, which issues a stream of new releases for each of several product lines – for example, word processor, spread sheet, etc, or compact car, sedan, sport utility vehicle. In constructing this model we are able to build upon a rich tradition of project modeling in system dynamics [29, 30, 37-39]. Each manager (agent) is assigned to a team managing one of the projects (i.e. a release for one of the software products). Each management team determines certain policies for the underlying system dynamics model. In our current simulation environment, the success of each project is evaluated at the end of each project cycle. The managers receive promotions and demotions according to the relative success of the projects on which they worked. The managers are then mixed and reassigned to new teams, in which each manager has an opportunity to learn from a team member. Learning occurs via a process of policy recombination where the probability of learning from a particular colleague increases with that colleague’s relative position in the management hierarchy (essentially, this is a genetic algorithm). The system dynamics model then simulates each project based on its managing team’s policy. Performance of each team is evaluated; promotions and demotions are handed out; and the cycle begins anew. This initial simulation environment has allowed us a preliminary examination of the impacts of team vs. individual evaluation, number of teams, and team size on evolutionary efficiency. For example: Early results suggest that, when done correctly, team-evaluation performs almost as well as individual-evaluation in terms of being able to “point” to people who perform well. Further, as the number of teams increase, the likelihood of converging on an optimum policy increases, while the speed of convergence decreases [14, 17]. Finally, Team size appears to have at most only a small impact on organizational evolution. Simulator specifics: Our proof-of-concept simulation environment has been developed as an object-oriented program [40-42]. A schematic of the class structure is shown in Figure 2. The simulations are parameterized using a control panel as shown in Figure 3. The user can establish the number of teams, the learning profiles, and the range of policy values that can be established (chromLength). The number of teams impacts the number of generations required to see policy convergence, and thus the user can control the number and duration of the generations. Altering the start time and the integration time step (DT) controls the system dynamics simulation. Increasing the time step decreases the resolution of the model behavior. Finally, the user can choose the probability of learning (recombination) and the probability of innovations (mutations). We are also able to change the learning process by running different types of company profiles. The individuals in the company can work in teams of varying sizes or can work alone. Performance in the company can also be rewarded by different promotion schemes. The company profiles are coded into the simulator. 5 Teams must be randomly reformed periodically for this to work. Essentially, the time series of how an individual’s teams have performed provides enough resolution to estimate that person’s average contribution. Code to write Correct code Undiscovered bugs WritingCode Writing code correctly