Information Network Topologies for Enhanced Local Adaptive Management

Abstract

We examined the principal effects of different information network topologies for local adaptive management of natural resources. We used computerized agents with adaptive decision algorithms with the following three fundamental constraints: (1) Complete understanding of the processes maintaining the natural resource can never be achieved, (2) agents can only learn by experimentation and information sharing, and (3) memory is limited. The agents were given the task to manage a system that had two states: one that provided high utility returns (desired) and one that provided low returns (undesired). In addition, the threshold between the states was close to the optimal return of the desired state. We found that networks of low to moderate link densities significantly increased the resilience of the utility returns. Networks of high link densities contributed to highly synchronized behavior among the agents, which caused occasional large-scale ecological crises between periods of stable and high utility returns. A constructed network involving a small set of experimenting agents was capable of combining high utility returns with high resilience, conforming to theories underlying the concept of adaptive comanagement. We conclude that (1) the ability to manage for resilience (i.e., to stay clear of the threshold leading to the undesired state as well as the ability to re-enter the desired state following a collapse) resides in the network structure and (2) in a coupled social–ecological system, the systemwide state transition occurs not because the ecological system flips into the undesired state, but because managers lose their capacity to reorganize back to the desired state.

Appendices

Appendix 1

Table 1 Appendix 1: Parameters

Appendix 2: Explanation of the Differential Equations

Equations 1 and 2 are taken from a standard textbook in theoretical ecology (Case 1999). The first term in Equation 1 describes the growth of crop that is limited by the carrying capacity (K). The second term is the amount of crop that is eaten by pests, and it is assumed to follow a type II function response relationship. The first term in Equation 2 is the same as the second term in Equation 1, except for the metabolic conversion rate M _r , which defines how much of the eaten crop is converted to pest biomass. The second term is the invasion of pests from surrounding areas. The third term is pest mortality and the fourth term represents the amount of pest that is eaten by birds. Bird predation on pests is assumed to follow a type II function response relationship.

Appendix 3

Table 2 Appendix 3: Simulation results

Appendix 4: Analysis of Simulation Variations

The objective of this study was to capture general characteristics of the model’s behavior. Hence, no in-depth statistical analysis was conducted. A number of experiments were, however, carried out to determine a suitable number of runs at appropriate time spans to produce fairly stable results (i.e., by reducing individual variations between simulation runs). Figures A1 and A2 illustrate the different variations. The stable behavior illustrated in these figures was kept for all other configurations (e.g., link densities, external variations, and scenarios).

Figure A1

Harvest distribution for low-link-density networks (3.2 links per node) and low external variations. The solid line is the mean taken from 10 different runs, and the dotted lines are the individual runs. The simulation duration was set to 4300 time steps (i.e., years), where the harvests below 300 time steps were discarded. As seen, individual variations are fairly low, each capturing the general characteristics of the simulation. The error bars represent 95% statistical significance (taken from all 10 runs).

Figure A2

The distribution of collapse durations with low-link-density networks (3.2 links per node) and low external variations. The solid line is the mean taken from 10 different runs, and the dotted lines are the individual runs. The simulation duration was set to 4300 time steps (i.e., years), where the first 300 time steps were discarded. As in Figure 9, individual variations are kept fairly low and each run captures the general characteristics of the simulation. The error bars represent 95% statistical significance (taken from all 10 runs).