Calibration of the Research Processes

Unlike the agent characteristics, which are initialised with empirical firm data, a set of free system parameters governs the model processes. To calibrate the model, these parameters have to be adjusted in such a way that the simulation output renders the empirical data of the real-world reference system (see Fig. 2 and Table 1).

The particular empirical focus of our modelling approach imposes empirically justified restrictions on the calibration process. To this end, a two-step calibration strategy is followed, called fractional factorial design (Thiele et al. 2014). In the first calibration step, representing the empirical restriction, a range of possible values for each parameter is defined based on expert knowledge. In the second calibration step, the search for the best fit between simulated and observed output data is performed by systematic parameter sweeping in the preselected part of the parameter space. In our case, two patent-related quantities are chosen as matching criteria between empirical and simulated data: (1) the total number of patents in the firm population and (2) the patenting profile of the population, i.e. the distribution of these patents over the patent classes.

The empirical reference dataset is the patent performance of Austrian biotechnology firms in the 5-year period of 2008-2012, and the calibration is performed with the simulated patent output after 20 time-steps (four time-steps representing 1 year). The best fitting parameter values are chosen in such a way that the empirical and simulated patents exactly match in terms of their total number (criterion 1), and reach the highest possible similarity with respect to their profiles, whereby the similarity of the profiles is measured as the Pearson correlation coefficient of the corresponding vectors in the knowledge space.

Table 1 Calibrated system parameters





Gridlock (agr‘)

Agent share with no research


Conservative (acon)

Agent share with conservative search strategy


Incremental (anc)

Agent share with incremental search strategy


Radical (arad)

Agent share with radical search strategy


Search dispersion (rsd)

Search radius for technology classes


Spillover (psp)

Probability of local knowledge spillover


Coop (aco)

Share of agents conducting cooperative research


Internal (a‘nt)

Share of agents conducting internal research


Success rate coop (srco)

Probability of successful cooperative research


Success rate internal (srint)

Probability of successful internal research


Patenting rate ( p0)

Base patent probability


Note: By definition, the parameters referring to population shares have to fulfil the constraints agri + acon + ainc + arad 1 and aco + aint = 1

Note that by choosing this empirical calibration strategy, one implicitly refers to the actual institutional framework conditions the Austrian biotechnology industry was embedded in, including the latent R&D policy during the calibration period. Thus, this parameter set defines the baseline scenario, as a reference for the simulations presented in Sect. 5.

< Prev   CONTENTS   Source   Next >