Abstract
Recent research on maintaining diversity in parallel problem solving takes into consideration only network structure, without considering the agents’ learning strategies. In this paper, we use a simulation study to extend March’s classic model by using locomotion and assessment as agents’ problem-solving strategies. First, we present a simulation framework that consists of external environment, communication networks, and agents’ learning strategies. Second, based on the framework, we develop March’s model to depict external environment. Third, we introduce four archetypical networks: a regular network, a small-world network, a preferentially attached network, and a totally connected network as agents’ communication structure. Finally, we design three experiments to explore the performance implication of locomotors and assessors under different networks. Results suggest that network structure affects performance more than learning strategy. The more efficient the network is at diffusing knowledge, the better the performance in the short run but the worse in the long run. Locomotors can help keep diversity; a high proportion of locomotors’ team has a better final performance but need more equilibrium time. Furthermore, moderate composition among locomotors and assessors increases costly interaction uncertainty. We discuss the findings’ implications for the regulatory mode and problem-solving literature.
1. Introduction
Much of practical problems consist of a “large number of parts that interact in a nonsimple way” [1]; such complex problem solving can be conceptualized as a parallel problem-solving process, which is the popular phenomenon in social science and engineering. A useful metaphor of a “rugged performance landscape” is introduced by Levinthal and Siggelkow to illustrate the complex problems and search process [2–4]. Problem complex arises from choices spanning over the horizontal dimensions and each possible combination of choices is represented as a performance score along the vertical dimension. The performance landscape is full of peaks meaning the higher score and valleys representing the lower score. As boundedly rational actors, agents are randomly located on the landscape [5, 6]. They interact with each other and hold different learning strategy to improve the team performance.
Facing parallel problem agents are hampered by a myopic search process. Since most of these peaks are “local” peaks, one that is easily leading the agents stuck into the local optima. Although current computational modeling research on problem solving guides us to how to assess the nature of the fitness landscape, two key factors need to be addressed in order to generate more meaningful contributions: network structure and learning strategy. The core part in understanding parallel problem solving is how team members solve problems collectively. Social network literature assumes that agents gain information from their communication network. We assume that agents will explore the problem space meanwhile watching the others [7]. Agents look at what other agents are doing. If they find a superior solution from their neighbors with who are connected, they will imitate that solution. Network structure, therefore, determines how well the knowledge about the solution is diffused in the team. In parallel problem solving, communication network affects the collective performance. Furthermore, although several studies have examined the outcomes of agents’ cognitive representation, they have paid less attention to locomotion and assessment, two popular orientations towards movement [8]. According to regulatory mode theory, we define locomotion as shifts from state to state and assessment as comparison between options [9]. Overall, consistent with extant literature, we move beyond a simple rugged performance landscape, since such traditional model cannot take network structure and behavior orientations into account. We, therefore, propose a computational model linking network structure and locomotion and assessment orientations together, which could then shed light on how to maintain diversity in parallel problem-solving research.
The main contributions of this paper are summarized as follows:(1)To the best of our knowledge, ours is the first study to simulate locomotion and assessment orientations in a computational model to explore the nonlinear relationship between search behavior and dynamic environment, extending the traditional March’s model.(2)We extend the current research on regulatory mode theory in the context of interaction pattern. Most studies on regulatory mode are rooted in psychology. We highlight its value in predicting agents’ problem-solving strategy and believe this may shed light on artificial intelligence research.
The rest of the paper is organized as follows. “Related Work” introduces the related work about NK model, March’s model, and their variants. “Model” proposes the definition and characteristics of performance landscape, searching strategy, and our algorithm. The experimental results are reported and analyzed in “Result” part. “Conclusion” part concludes the performance implications of search strategies on changing performance landscapes and discusses the future work.
2. Related Work
2.1. Exploration versus Exploitation in Parallel Problem Solving
Imagine the situation of a team leader who takes in charge of the whole software development process, during which all the team members face the complex problem and share the new knowledge to solve the problem. This scenario is what we label parallel problem solving, in which all the team members face the same complex problem, and any agent who gets progress will not affect others [10]. The payoff of any agent’s achievement is independent of others’. Besides, agents can observe other agents’ action, absorb and learn the successful solution, and explore the next new possible alternatives based on the previous one. Facing the parallel problem solving, there are two challenges confronting the team leader: one is how to organize the interpersonal communication structure among the team members. According to the previous literature, network structure influences the learning speed, and there is a significant tradeoff between learning speed and average performance. The other one is to how to control the ratio between assessment and locomotion, which are popular cognitive behavior model among team members. Psychologists have paid considerable attention to the consequences of locomotion versus assessment, which definitely affect the learning speed and affect the final performance. In a word, the team leader should decide what kind of communication structure is applied. Is it better to let team members involve in a density network, a proven successful solution will spread fast among the team.
The balance between learning speed with final performance can be seen as the trade-off between exploration and exploitation [11, 12]. According to the seminal research of James March, exploration can be viewed as introduction of new information and development of new solutions, while exploitation involves developing existing knowledge [13]. In March’s simulation model, there are fast learners and slow learners. Fast learners exploit the organization’s existing knowledge, which leads to a high-performance level in the short run. Fast learners, however, are likely to reduce the diversity of organizational solutions, which leads to a low performance level in the long run. Slow learners, though less efficient, allowed the firm to keep diversity of agents’ solutions, which enables the team to explore a higher-level performance of possible combinations of solutions and leads the team to an optimal equilibrium. March’s work demonstrates the trade-off between exploration and exploitation, which represents a fundamental conflict between short-run and long-run development in almost all complex self-adaptive systems. In Lazer and Friedman’s computational work, parallel problem solving can also be seen as a balancing of exploration and exploitation [14]. In parallel problem solving, agents can explore the possibility of combination of solutions or exploit the existing solutions by imitating the neighbor’s solution through interpersonal network.
2.2. Network Structure
Agents will learn each other’s solution through interpersonal network. In March’s model, an organizational code was designed to represent a hub-like structure in which agents’ superior solutions are put together, and agents, in turn, can learn from the organizational code. Although there seems no communication structure in March’s model, March in fact implicitly assumed a hub-like interpersonal network. Miller et al. extended March’s model by adding direct interpersonal learning, in which agents are randomly situated within a grid [15]. Agents can contact their direct four neighbors and learn from the best one. If the neighbors are inferior to the focal agent, agent then engages in “distant learning” by randomly drawing any other four agents from the whole organization and choose the best one to imitate. Believing that network structure will amplify the trade-off between exploration and exploitation, Lazer demonstrated the impact on the final performance level of networks: a linear network, a totally connected network, a variety of random networks, and a variety of small-world networks [14]. The results show that an efficiency network, which is very good at disseminating information and promoting agents imitate each other, will reduce the diversity in team and lead to a higher performance level in the short run but lower level in the long run, while inefficient network maintains diversity and leads to a higher performance level in the long run. Fang et al. follow the exploration and exploitation research stream and propose a very special network structure: semi-isolated subgroups [16]. In their simulation mode, organization is divided into some subgroups and there are some hub-like actors that connect the subgroups. Then, they explore how the degree of subgroup isolation and intergroup connectivity affects the trade-off between exploration and exploitation. Their model demonstrates that moderate levels of intergroup links will maintain the diversity and perform best. Particularly, Fang’s model assumes that exploration happens in each subgroup, in which agents conduct parallel problem solving and are isolated from the other subgroups, while exploitation happens in intergroup, in which agents learn superior solution across the group. In this way, a productive balance between exploration and exploitation will be achieved.
As we noted above, prior research has found that parallel problem solving involves the balance of exploration and exploitation and that communication network structure may amplify the problem by controlling the learning speed among agents [17–19]. Efficient network may reduce the diversity and destroy the possibility of future exploration. Some special network like semi-isolated subgroup may help to solve this problem.
2.3. Assessment versus Locomotion in Learning
Although many scholars admit that network structure plays an important role in balancing the trade-off of exploration and exploitation, little attention has been paid to agents’ learning strategies, which fundamentally affect learning rate and maintaining diversity. Prior studies following the March’s milestone work view the innovation and imitation as the basic behavior rules among agents [20, 21]. Each step agent either explores another new combination of knowledge element (innovation) or imitates other’s solution directly or indirectly. Actually, in practice, there is a wide range of behaviors; also psychologists have paid considerable attention to the behavior type under knowledge diffusion situation. It is certainly useful to consider more useful behavior rules when we simulate the team process.
Recent research has focused on agents’ strategies as a mechanism for problem solving. Wu’s seminal research proposed dummy queries to cover up user queries and thus protect user privacy [22–24]. Shigen et al. established the malware propagation model in WSNs according to the mixed Nash equilibrium strategy [25]. Zhou et al. proposed a Heterogeneous Susceptible-Infectious-Removed-Dead (HSIRD) model based on epidemiology to represent the HSN communication connectivity [26]. In a word, agents’ strategies are, indeed, conducive to performance.
According to regulatory mode theory, locomotion and assessment modes are considered to be both more general and more independent behavior rules [8, 9]. Regulatory mode theory proposes that locomotion involves movement from state to state. Traditional mode theory believes that locomotion refers to moving from the current state to a desired or valuable end, but recent studies show that the destination is unnecessary desirable or valuable, a locomotion can be any change of position. A locomotion prefers “move” to “stay”; even the current state is proven to be inherently positive. Most generally, “assessment” refers to comparing the current state and the end state. The keywords for assessment are measuring, interpreting, evaluating, and making comparisons.
In parallel problem-solving context, as the essential nature of assessment involves comparing oneself to some norms, agents with a high assessment orientation are sensitive to social standards and they want to improve their performance. High assessors are more likely to conduct long-distance learning, that is, considering the whole alternatives in team and choosing the best one to imitate. By contrast, as the essential nature of locomotion involves moving from one state to another without particular destination, agents with a high locomotion orientation intent to do “something, anything, other than the same thing.” High locomotors are more likely to explore any possible combination of knowledge element by themselves or imitate the direct neighbors’ solution. Although this movement may lead to an inferior solution, the movement itself satisfies the locomotors’ desire.
3. Model
3.1. Simulation Framework
In this section, we describe the simulation experiment design and model draw on March’s work. In Figure 1, we show the simulation framework adopted by this paper as well as examples (where the green profile represents the locomotors while the red represent the assessors). From Figure 1, we can see that simulation model consists of external environment, communication networks, and agents’ learning strategies, which will be described, respectively, as follows. Besides, three experiments are set to explore the relationship between agents’ search behavior and dynamic environment.
We regard an agents’ team as a complex adaptive system facing complex problem, where agents interact with one another based on communication network. In particular, we view agents as carriers of ideas and knowledge and team performance as a result that emerges from interactions among agents. Based on self-regulation theory, agents are either assessors or locomotors distinguished by different learning strategy. This assumption of learning strategy makes our paper distinct from March’s (1991) work. In experiment 1 and experiment 2, we will explore the performance implication of communication network with the same learning strategy, which means all the team members are locomotors or assessors. In experiment 3, team consists of both locomotors and assessors. Particularly, the proportion of locomotors and assessors is designed as 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1. We will explore the performance implication of these different proportions under different network structures.
3.2. External Environment Landscapes
To simulate the complex problem solving, prior literature uses NK landscape or March’s model to model the degree to which alternative choices are correlated with one another. In NK model, N represents the number of distinct choices in an overall solution. The variable K refers to the extent to which the payoff of one choice depends on others [27, 28]. The landscape is the mapping from N choices to a payoff value. The variable N definitely determines the complexity of landscape, while the variable controls the degree of rugged landscape (see Figure 2). Similarly, March used m-dimensional vector to model the external environment, in which every dimension is randomly assigned the value of 1 or −1. The bigger m is, the more rugged the external environment landscape is.
(a)
(b)
Both models can perfectly depict the complex environment and parallel problem solving. For instance, a number of decisions constitute the solution of a team’s software develop process, including the decisions about which programming language should be chosen, Java or Python, which jobs should be done in team or outsourcing. Notice that the payoff of some particular choice is correlated with others; for example, team members are more likely to use Python rather than Java, so if team leaders decide to do themselves, then Python will be preferred. Consider the situation agents will learn each other from interpersonal network, and Miller et al.’s extending work has been demonstrated successfully, we adopt March’s classic model [15].
Agents have their own knowledge about the external environment; their beliefs about each of the M dimensions can be the value 1, 0, or −1, which reflect the right, absent, and wrong. All the agents can adjust their knowledge through their own learning strategies, which determine the extent to which dimensions of the beliefs match each other.
3.3. Network Structure
Implicit in March’s model is the assumption that agents learn from an organizational code, which reflects the most beliefs among better performers. By contrast, Miller extended March’s model by introducing the interpersonal networks. Agents can learn from each other and it is unnecessary to exchange knowledge mediated by organization codes. Following the prior literature, in our simulation model, we examined four archetypical networks: a regular network, a preferentially attached network, a variety of small-world networks, and a totally connected network.
A regular network is one in which each node connects to its four direct neighbors (see Figure 3(a)). The regular network likes grid in Miller.’s research mentioned before. To hold the density constant, we use the same number of links to construct a small-world network. A small-world network is constructed based on Watts and Strogatz’s model. Both high cluster parameter and short path distances can be seen in a small-world network. To keep the same number of network ties with regular network, we then randomly cut down some local links and rewire the long-distance nodes. In this way, as Figure 3(b) demonstrates, the average path distance in regular network will be lowered. If we rewire more local links, the average path distance would drop rapidly. In the meantime, fixed number of neighbors makes sure the high cluster parameter, which makes the regular network into a small-world network.
(a)
(b)
(c)
(d)
A preferentially attached network reflects Matthew’s effect, which indicates that the more friends you already have, the more new friends you will have. New links preferentially attached to the nodes whose links are the most. In our model, every agent links to another at a certain proportion; occasionally agent may become isolated vertex with very low probability (see Figure 3(c)).
A totally connected network (see Figure 3(d)) is one in which every node connects with every other node. In simulation, we keep the number of nodes constant, and then construct a totally connected network.
Because of the same number of network ties and nodes, regular network and small-world network have the same density which is defined as the potential ties divided by the real existing ties. Obviously, the density of a totally connected network is the maxim, 1. The average path distance, which is defined as the steps one node takes to reach another one, is quite different among the three networks. In simulation, we will explore the impact of density and average path distance on final performance.
3.4. Learning Strategies
Having introduced the complex network properties of the model, we will turn our attention to agents’ learning strategies, accounting for regulatory mode theory locomotion refers to moving from one state to another. Assessment refers to measurement, comparison, and evaluation. Locomotors keep moving while assessors love evaluating. A proportion q (0 < q < 1) of the beliefs is constructed to reflect the difference between the two agents.
Agents learn through complex interpersonal network. In each period, agents involve finding the best performers among their neighbors who are directly connected. Once best performers are identified, agent will update q proportion of the beliefs to that superior neighbor. Consider the characteristic of locomotion and assessment we discussed above, locomotors have low value of q proportion because of their intent to move. Locomotors cannot wait for a long evaluation time to move from one state to another; they are eager to change the status quo no matter what this change will lead to their future performance. On the contrary, assessors love comparing the details between their beliefs to the superior neighbors. A high value of q proportion of assessor’s beliefs will be updated according to their superior neighbors.
Another two situations should be taken into our consideration: if all the neighbors have the equal best beliefs, then locomotors or assessors will choose one randomly as the target to imitate. If all the neighbors have beliefs inferior to the agents, then the agents stop updating their beliefs. Implicit in this design of learning strategies is the assumption that agents can learn from each other in their own way. It is reasonable to assume that different learning strategies lead to different learning rate, regardless of the structure of the interpersonal network. Even though the interpersonal network’s density may be the same, the aggregate team-level performance differs depending on the composition of assessors and locomotors. Implication in design of the different imitation proportion q of assessors and locomotors is the assumption the agents differ in their frequency with which they update beliefs. We incorporate this assumption into the model by having considering the agents different learning strategies. These output differences are reflected in the speed the team gets equilibrium and the final performance achieved.
4. Results
4.1. Implementation
We use MATLAB R2011b to construct our model and its picture software toolkit to report the final result figures. We consider a team of 50 agents who engage in solving a parallel problem. The problem is modeled as an m-dimensional vector whose element is assigned 1 or −1 randomly. Agents hold the same m-dimensional vector whose element is assigned 1, −1, or 0, respectively, meaning the right, the wrong, and not sure with the reality. The agents’ initial beliefs are randomly generated, and those 50 agents are randomly placed on complex interpersonal network. Four networks are designed following the algorithm discussed above. Every agent’s performance is calculated as the proportion of their right beliefs. In each period t = 1, 2, 3, … , T, all agents seek the best performance of neighbors and update their beliefs with a certain probability, and we tracked the average performance of the team. Through repeated attempts and experiments, we put 200 as the equilibrium time when all the agents had the same performance levels. We adjusted the parameter values of the model and let every simulation model run for 100 times and calculated the average performance. The detailed summary of model parameters can be seen in Table 1.
4.2. Analysis
4.2.1. Experiment 1: Locomotion Learning Strategy on Different Network Structures
Figure 4 shows the performance over 100 simulations of four interpersonal networks under locomotion learning strategy over time. When all the agents are locomotors, Figure 3 demonstrates clearly the four patterns: the totally connected network is the first one to find the best solution and outperforms the other networks in the short run, but in the long run, regular network and small-world network perform better than it. Because of the isolate nodes, preferentially attached network cannot be a single component, which leads to the worst performance level.
When we focus on three single-component network structure: regular network, small-world network, and totally connected network, we can get the conclusion that having fewer communication opportunities will improve the long-run performance. Our model is close enough to the prior literature, March’s model and Miller’s model. The fully connected network drives out diversity with a very fast speed, which leads to a local optimum. On the contrary, regular network and small-world network can keep the diversity for quite a long time, which means that most agents’ beliefs are worse than the best one in team, but a better solution will be found in the long run. The diversity will allow the regular network and small-world network to improve the solution in the long run and finally reach the global optimum. Small-world network is much more efficiency at diffusing knowledge due to the short average path even though they have the same density. When we compare the results of regular network and small-world network, we can conclude the more efficiency at diffusing information the lower performance level they reach in the long run.
4.2.2. Experiment 2: Assessment Learning Strategy in Different Network Structures
When all the team members are assessors, we can get four similar performance curves as Figure 5 shows. Figure 5 demonstrates a very clear trade-off between effective at information diffusing and final performance level. Fully connected network has the maximum density and small-world network has the shorter average path, which let all the agents quickly update their beliefs when they find the best neighbor. Furthermore, compared to locomotors, assessors are more influenced by their neighbors. Due to their strong concerns with comparing their own solution to the best neighbors’, assessors are more likely than locomotors to evaluate and imitate continuously to the best neighbor’s solution, which leads to a high value of proportion of m-dimensional vector changed. In a word, assessors are distinct from locomotors by higher self-evaluation concerns, which leads to more sensitivity to performance “standards.” Assessors are the best learners who try their best to imitate the target.
As we discussed above, the results in Figure 5 can be explained by the agents’ learning rate. Assessors learn faster than locomotors which makes the comparison of short-run performance and long-run performance on the four complex interpersonal networks much clearer. This “tack effect” makes fully connected network find the best solution in extremely quick time (less than 10 periods); all the assessors climbed uphill from the best solution that exists in team. Although regular network reaches the highest performance level in the long run, its convergence time is too short (no more than 40 periods). As we discussed above, as well as final performance level, the equilibrium time will be reported in the next section.
4.2.3. Experiment 3: Final Performance Level and Equilibrium Time in Four Complex Networks
The literature on knowledge sharing highlights the importance of effective knowledge diffusion, which is believed to lead to a high-performance level, while diversity is found to be beneficial to the final performance level. Figure 6 demonstrates these two positive and negative effects. Learning strategies and communication networks both affect knowledge diffusion and diversity. More specifically, fully connected network and assessment both positively affect knowledge diffusion, which facilitates the best solution exchanged, but negatively reduces diversity, which is also important for final performance level. Although regular network and locomotion are inefficient in diffusing the solution, they keep the diversity, which leads to a high final performance level.
Equilibrium time is a proxy for diversity. A high value of equilibrium time reflects the fact that it takes the agents a long time to climb the performance hill. Figure 6 shows that the combination of regular network with locomotors is the best way to keep diversity. Although both regular network and locomotion help keep diversity, their strength is quite different. Obviously, network has much more powerful strength to do that way. In a real team situation, controlling the frequency of communication in a team is a quite effective way to keep diversity. In fact, it is not always the case that all the team members are locomotors or assessors; we should consider the performance of different composition of the two leaning strategies.
4.2.4. Experiment 3: Equilibrium Time and Final Performance Level among Different Team Compositions
Prior literature on team composition from social, psychology, and organization is used to explain how team composition influences the team performance. According to the organization theory and complexity research, some common variables such as personal demographics, personality attributes, values, and other characteristics (KSAOs) are considered as the main mechanism to influence team performance. Generally speaking, deep-level composition variables, such as cognitive pattern and behavioral mode, and some psychological variables have a much stronger influence on performance than surface-level variables which can be knowledge, skill, or other personal demographics. It is not convenient or high cost to measure team member’s behavioral pattern; simulation research gives us a valuable opportunity to construct a virtual experiment to explore how team composition affects the performance.
Consider a team with both locomotors and assessors, how this different composition affects the equilibrium time and final performance level. We designed 10 teams with 10 percent to 100 percent locomotors. Figure 7 tells us that the equilibrium time in general will be long very significantly as the proportion of locomotors increases. Locomotors prefer to move to change the current state. For locomotors, they could accept performance decline as long as escape from the current state. A little change can satisfy locomotors’ motivation to move. A team composed of many locomotors will face a large innovative space because much more diversity will be maintained among team members, which leads to higher final performance level and longer equilibrium time (as Figure 8 shows).
What surprised us is that when the proportion of locomotors is 0.4, the team performance declines. When the proportion of locomotors is moderate, as 0.5 or 0.6 for example, the performance growth slows down. This result reflects the complexity of a self-adaptive system: when the proportion of locomotors or assessors is moderate, the interaction is quite complex and nonlinear, which leads to performance growth slow down or even decline. On the contrary, a majority or minority of locomotors or assessors shows a clear relationship between equilibrium time, final performance level, and team composition. Furthermore, Figure 7 again shows us that network configuration is the main power influencing the team performance notwithstanding the proportion of locomotors or assessors is large or small. Regular network, which can keep diversity best, will achieve a high-performance level all the time. No matter what composition the team is, the more efficient the team configuration at diffusing knowledge, the better the final performance in the long run and the worse in the short run.
5. Contribution and Discussion
In this paper, we extend Miller’s model by considering the locomotion and assessment and explore the performance implication of network structure and agents learning strategies. Given the current limited understanding of performance implication of agents’ learning strategies, we consider regulatory mode theory and introduce locomotion and assessment modes into our model to simulate the real team on parallel problem solving. Simulation results indicate that both network structure and agents’ learning strategies can affect team performance.
Our simulation results conclude both positive and negative effect of network structure on team performance. Efficient network is good at diffusing knowledge, which reduces diversity, which positively affects the short-run performance but negatively related to the long-run performance. Fully connected network is very efficient at disseminating knowledge; that is, once a solution is proven to be better, it will spread to the whole team simultaneously because all the agents are connected with each other, which leads to a high-performance level in the short run but a low-performance level in the long run, because only one solution exists among agents, cohesive network reduces the diversity, which can improve the performance in the long run. According to regular mode theory, in our model assessors are more likely than locomotors to imitate the best performed neighbors, which leads to a similar effect like network structure. The more assessors (or less locomotors) a team has, the shorter (or the longer) equilibrium time it will be, and the higher (or the lower) short-term performance and the lower (or the higher) long-term performance.
To balance the exploration and exploitation in team, managers should compose the team ahead of time to maintain team competitiveness. Members who hold different learning strategies should be selected to optimally fit each other. From the simulation results, we can conclude that assessors are good at exploiting current knowledge, which helps team survive in the short run, while for the team who has a majority of locomotors, they should realize their long run potential. Locomotors are slow learners who are preferred to maintain diversity. Team can facilitate future learning where a majority of locomotors exists. Furthermore, moderate composition among locomotors and assessors increases costly interaction uncertainty. The complexity of this moderate composition leads to performance growth slow down or even decline. This simulation results seem to argue for composing team under less complexity. Prior literature overlooks the role of complexity in maintaining the competitive strength. It is a great challenge to managers to configure the moderate locomotors or assessors.
There is huge potential for extending our simulation model. Some dimensions should be taken into the future consideration, such as exploring the implication of changing environment. While we construct model in a stable external environment, one may imagine some exogenous shocks which are episodic or smoothly continuously would alter the problem landscape. Does that reduce the value of locomotors learning in the long run, or does it help assessors to maintain diversity which yields competitiveness in the long run? Any of these dynamic factors would be reasonable to add to our model, such as turnover. Prior studies have demonstrated that turnover can be advantageous if the environment changing is turbulent. Likewise, if turnover is permitted, the dynamic proportion between locomotors and assessors would facilitate team adaptive ability to a changing environment. Furthermore, we hope our simulation results be supported in empirical research. In a real experiment or actual team, empirical investigation should be done to test our propositions and extend our validity.
Data Availability
We use simulation to build and test our model. All the codes are written by the authors. If you are interested in their coding, contact the corresponding author.
Conflicts of Interest
The authors declare that they have no conflicts of interest regarding the publication of this paper.
Acknowledgments
This study was supported by grants from the Ministry of Education in China “Project of Humanities and Social Sciences” (Project no. 17YJA630132), National Social Science Foundation (Project no. 21BGL054), and National Natural Science Foundation of China (Project no. 71802137).