Home Computer Science
The sets I, II and III (Table 4) allows us to compare the behaviour of highly specialized actors (s = 0.5) in a dynamic environment (v = 0.2) for three different values of the exploration capability (p = 0.1; p = 0.5; p = 0.9).
The sets IV, V and VI (Table 5) permit to analyze the impact of different values of the exploration capability of ICAs in a more static environment (v = 0.8) and for the same high specialization (s = 0.5).
The sets VII, VIII, and IX (Table 6) refer to a dynamic environment (v = 0.2), to a low specialization^ = 0.8) of ICAs populating the system and to respectively three different values of exploration capability (p = 0.1; p = 0.5; p = 0.9).
Finally, the sets X, XI and XII (Table 7) report the behaviour of a low specialized system (s = 0.8) in a static environment (v = 0.8) for three different values of the exploration capability of ICAs.
The simulations output variables are described in Table 8.
Results of Simulations
Assessing the internal coherence of the model implies to verify that the code is free of evident bugs, that it works coherently with the meta-model and that the agent- based computational model is able to reproduce some stylized representations characterizing the system under investigation. Through the simulation experiments we aim at exploring if and under which conditions the micro specifications we implemented in the agent-based model are able to produce some known regularities characterizing Regional Innovation Systems framed as complex networked learning systems.
In our experiments we analyze the behaviour of the system for different levels of the capacity of exploration (p) and different degrees of competences’ specialization under two competitive scenarios: a static and a dynamic one. Figure 5 reports the
Fig. 5 Number of surviving agents (s = 0.5)
Fig. 6 Number of surviving agents (s = 0.8)
number of surviving ICAs of a high specialized system (all the ICAs have the same level of scope s = 0.5) under the three different levels of the capacity of exploration (p) in a static and in a dynamic competitive environment. Figure 6 shows the trend of the same output variable (number of surviving ICAs) in the case of a low specialized system (s = 0.8).
The number of surviving agents in the system increases as their capacity of exploration increases, both in a static and in a dynamic environment. This is an expected and plausible result; in particular, this result is coherent with the seminal paper of March (1991) on exploration and exploitation in organizational learning systems and with most of literature on learning systems deriving from March’s original investigation.
In Figs. 5 and 6 we can also find other plausible results that support the verification of the implemented model. As showed in Fig. 5, for low (0.1) and medium (0.5) values of the parameter p the number of surviving ICAs is higher in static than in dynamic competitive environments. Furthermore, the difference between the trends of population referring respectively to static and dynamic environments is significantly reduced (Fig. 6) when the level of competences specialization in the system is low (for s = 0.8). It seems that less specialized ICAs are able to react better than more specialized agents to a high environmental volatility. Agents that cover a more wide range of required knowledge, as expected, show a higher capability to survive in dynamic environments and seem to be less affected by the change in the competitive Environment’s Regularity.
Fig. 7 Average number of frames for each ICA in a static environment (s = 0.5)
Fig. 8 Average number of frames for each ICA in a dynamic environment (s = 0.5)
Figures 7 and 8 add new elements to the above analysis. Figure 7 reports the average number of Frames of ICAs for each of the two IO produced during the simulation by a static competitive environment, under the three different levels of exploration capacity (p). The number of Frames in the individual memories of Competent Actors of the system can be interpreted as a proxy of the learning performance of ICAs. In a static environment, individual learning performance increases according to the individual capacity of exploration (better performances are achieved for medium and high values of p).
Figure 8 shows that in a dynamic environment (v = 0.2)—five different ER are sent to ICAs by the CE during the simulation time—better individual learning performances are achieved when the value of their capacity of exploration is equal to 0.5. High levels of exploration, in this case, produce the same result of low levels of exploration.
This result supports the verification of the code and the coherence between the implemented model and the meta-model (Sect. 5). In fact, we know from the literature (Benner and Tushman 2002; Gupta et al. 2006; March 1991) that in a high volatile environment high levels of exploration produce high exploration costs that are not sufficiently remunerated by adequate rewards (the time needed to complete a cycle of exploration-exploitation is longer than a market cycle).
This result highlights the need for additional experimentation and investigation, as the literature claims for a balance between exploration and exploitation in order to sustain organizational and system performances in the long run. In the model proposed here, the ambidexterity mechanism (exploration and exploitation activities are performed at the same time and are balanced through organizational and collaborative structures) of balance between exploration and exploitation (Benner and Tushman 2002; Gupta et al. 2006) is implemented but not sufficiently articulated in its implications. The activity of exploitation is modelled as partnership creation to produce Collective Interpretations. This capability, in the actual version of the model, depends on several parameters (the competence c of the agents, the propensity toward collaboration T, and the Hamming distance between two Individual Interpretations). At this stage of the research the impact of these latter parameters has not been investigated through specific experiments.
Finally, the comparison between Figs. 7 and 8 outlines the need for additional experiments and analyses useful for theory building. One element of coherence with the literature on exploration and exploitation in learning systems can be detected, in particular in Fig. 8. As suggested by March (1991), in highly complex environments the learning process is a sort of random walk. Learning performances are strongly affected by randomness. In our case, high volatility (v = 0.2) is interpreted as high complexity. Better performances are achieved when the capability of exploration of all ICAs of the system is equal to 0.5. p = 0.5 means, from a computational point of view, that the strategy of exploration of agents in the system is a random strategy, resulting, in high dynamic and complex environments, in a more efficient search and solution strategy and in better learning performances.
Figures 9 and 10 show the distribution of the collective interpretations’ dimensions (the number of individual Frames contributing to the collective
Fig. 9 Collective interpretation’s dimension (dynamic environment, s = 0.5)
Fig. 10 Collective interpretation’s dimension (dynamic environment, s = 0.8) interpretations) with respect to the levels of the capability of exploration and respectively for high (Fig. 9) and low levels (Fig. 10) of competences specialization (the behaviour of the system is less affected, in this case, by the volatility of the CE). By comparing the two figures we can identify an expected and plausible result that again confirms the existence of an internal coherence of the implemented code. In fact, the dimension of Collective Interpretations is in general greater in the case of highly specialized systems than in low specialized ones. In highly specialized systems agents are characterized by a low capability to cover the spectrum of competences required by the CE. Thus they need to create links with other agents in order to build up a more complete Interpretation. This diffused trend toward networking results in the presence of networks, which, in some cases, are composed of more than 9 agents. This latter case is not shown for low levels of specialization. As said before, this result supports mainly the test of the code and does not add any significant element to theory building.
Finally, Figs. 11 and 12 show the behaviour of the system in terms of economic performance, with respect to the volatility of the environment and under different combinations of competences’ specialization and exploration capability of ICAs.
In particular, Fig. 11 reports the trend of the economic performance (measured as the difference between the capital assigned to the system at the beginning of simulation and the capital registered at the end) of a high specialized system (s = 0.5) with respect to the three levels of the capability of exploration (p). Figure 12 refers to the trend of economic performance of a low specialized system.
Fig. 11 Mean delta capital in the system (s = 0.5)
Fig. 12 Mean delta capital in the system (s = 0.8)
Table 9 Main results of internal verification
As expected and according with the design of the agent-based model, better performances are achieved on average in static environments. Indeed, in a static environment the need for ICAs to sustain costs of exploration and exploitation are reduced with respect to the case of a dynamic environment, in which the messages sent by the CE to the ICAs are more frequently changed. Furthermore, according with results on the population dynamics showed in Figs. 5 and 6, better economic performances are achieved in those systems populated by more explorative ICAs (high levels of the capability of exploration). Finally, the comparison between Figs. 11 and 12 highlights that higher levels of economic values are obtained in low specialized systems (s = 0.8). This latter result is again coherent with results reported in Figs. 5 and 6.
Table 9 shows a brief summary of the main results of the internal validation. Finally, we can conclude that: