Model Examination and Application

The model is implemented in Java using the agent-based modelling platform MASON[1] and the results are analysed with R. The simulations are conducted over 120 time steps (i.e. a period of 30 years), and the results represent averages of 100 runs with varying random seeds. This section deals with model testing and application to different scenarios that can be interpreted as relevant policy scenarios in Austria. In each of the scenarios, the technology profiles of the agent population that emerge in the long run are compared and interpreted.

Robustness Tests

An essential step for the credibility and validity of the model is testing the model’s robustness against the randomized initial conditions (described in Sect. 4.1). One way to easily generate robust results is averaging over a high number of simulation runs to smooth the effects of possible outliers. Hence, this subsection provides insights into the model behaviour through the analysis of 100 single runs for the baseline scenario given by the set of system parameters in Table 1, and for interesting alternative scenarios with changed system parameters.

To this end, two scenario pairs are formed for testing parameter variations with respect to the baseline scenario. The first pair is aimed at analysing the effects of altered cooperative behaviour of the agents (Table 3). On the one hand, the share of cooperative agents, the spillover rate and the probability of successful cooperative research are increased (Scenario 1a). On the other hand, the share of cooperative agents is reduced as well as the probability of knowledge spillover (Scenario 1b).

Each scenario was analysed regarding the total number of patents after 120 simulation steps (i.e. 30 years). The distribution of the patent counts over 100 runs for the scenarios 1a and 1b and the baseline scenario are illustrated as estimated Gaussian kernel density functions in Fig. 3. Evidently, the number of obtained patents is the highest for the scenario of increased cooperative activities. The patent counts for the opposite scenario and the baseline scenario are overlapping but nevertheless, the median value of the baseline scenario exceeds the one from the scenario with reduced cooperative research. The values for the variances indicate lowest variability in the reduced cooperation scenario, and the highest variability in the increased cooperation scenario (almost twice as high as the baseline scenario). However, there is no linear correlation between the number of patents and the variability of the scenarios.

The second scenario pair refers to the prevalence of research strategies in the agent population. Scenario 2a is characterised by an increased share of agents with radical search strategy in combination with an increasing search dispersion

Table 3 Scenarios with different degree of cooperation

Scenario

description

Coop (aco)

Share of agents conducting cooperative research

Spillover (psp) Probability of local knowledge spillover

Success rate coop (srco) Probability of successful cooperative research

Scenario 1a

Scenario 1b

#

#

Note: ", increased value with respect to baseline scenario; #, decreased value with respect to baseline scenario; !, unchanged value compared with baseline scenario

Distribution of patent counts over 100 runs (baseline and scenario pair 1)

Fig. 3 Distribution of patent counts over 100 runs (baseline and scenario pair 1)

parameter (regulating the search radius for the agents while searching for new technology classes). For scenario 2b, the shares of agents with conservative and incremental search strategies are increased along with a reduction of radical agents.

Again, the kernel density functions are plotted to visualise the distributions of the 100 different runs for the scenarios (Fig. 4). The three scenarios clearly differ with respect to their total number of patents. It becomes evident from the values of the variances that the higher the numbers of obtained patents, the higher is the variability of the scenario. What all functions have in common is a kink to the left of the median: intuitively, the higher the patent counts of the respective scenario, the

Distribution of patent counts over 100 runs (baseline and scenario pair 2)

Fig. 4 Distribution of patent counts over 100 runs (baseline and scenario pair 2)

Table 4 Scenarios with different prevalence of research strategies

Scenario

description

Radical (arad) Agent share with radical search strategy

Conservative (acon) Agent share with conservative search strategy

Incremental (amc) Agent share with incremental search strategy

Search

dispersion (rsd) Search radius for technology classes

Scenario

2a

"

Scenario

2b

#

"

Note: ", increased value with respect to baseline scenario; #, decreased value with respect to baseline scenario; !, unchanged value compared with baseline scenario more pronounced the kink. This indicates that there are generally a few runs with low patent counts below the median, whereas the majority of the observations is scattered around the median. However, this finding is not so distinct in the first scenario pair displayed in Fig. 3. There, it even is the case that there is a slight kink to the right of the median.

  • [1] http://cs.gmu.edu/~eclab/projects/mason/
 
Source
< Prev   CONTENTS   Source   Next >