5. Conclusion and Implications
5.5 Limitations
This section discusses two limitations to the AIE model, the computing requirements and the lack of evolutionary processes such as learning. Section 5.5.1 discusses the computing requirements issue and section 5.5.2 discusses the evolutionary and learning issues.
5.5.1 Computing Power
Ten years of computer processor time were used in developing the AIE model for this thesis, which includes time for numerous detours and incomplete jobs. However, it does serve to indicate the serious amount of computing power required to calibrate the AIE model. This limitation could be address in several ways.
Since the timeliness of predictions is usually an issue, a highly concurrent system of 121 processors to calibrate the 121 network topology models would be advantageous. Currently this amount of computing power is available only to larger institutions; however this may not be the case in 10 years from now. If Moore’s Law holds true that the number of transistors to be placed onto an integrated circuit cheaply doubles every two years, the current quad processor desktop computers will be replaced with a 128 processor desktop computer in ten years.
An alternative approach is programming graphic adaptors of computers, which are massively parallel, to provide the potential to run all 128 network topologies simultaneously on one computer. This parallel programming approach is gaining ground and looks promising, as the amount of RAM in graphics cards increases and concurrent programming languages becomes more ‘user-friendly’. Currently, the most prominent instance of this is NVidea's CUDA system (http://www.nvidia.com/object/cuda_home.html#), which enables up to 512 cores to run simultaneously on a single PC. Distributed processing is also supported in Mathematica and many other numeric programs these days. Section 5.6.13 discusses further research into using graphic adaptors to decrease the time to calibrate the AIE model.
People donate CPU time on their home PCs for projects such as the Search for Extra-Terrestrial Intelligence (SETI) or prime number work. AIE could be run in a similar way by distributing the processes over a number of CPUs. However, establishing such a regime requires considerable coding effort.
AIE is coded in NetLogo a fourth generation language, which is itself written in Java. Java is a flexible interpreted language, which will run on almost combination of hardware and operating system. Writing the code in a third generation language such as C or C++ would provide a significant decrease in the time to calibrate the AIE model. Unfortunately development time is far greater, when using a third generation language than NetLogo, and requires more technical expertise.
Additionally, the calibration from the previous prediction can be used as the starting point for the search to calibrate the model for the next prediction, which saves considerable time. Another option is improvements in search techniques to calibrate the model, which could further reduce the search time and section 5.6.1 discusses this point in further research.
5.5.2 Learning and other evolutionary processes
AIE is limited with regard to model learning unlike evolutionary economics. This is a concern because firms or people do learn from experience and change the rules under which they operate. However, there are three considerations that ameliorate this concern. First, the AIE model has a short calibration and prediction period. Second, existing near-zero-intelligent ABMs capture the most salient feature of mature systems. Third, the firms in the D&B survey exhibit impediments to learning the true state of the world. This section discusses the three considerations.
First, the AIE model is calibrated over a short period, after the profit expectations undergoes a structural change or phase transition. If the phase change or structural change in expectations is considered a period of rapid learning or changing of rules, the period outside the phase change is seen a time of stable rules and slow learning where learning serves to reinforce the existing rules or provides insufficient impetus for entities to change their own rules. Then the calibration and prediction of the AIE model during a period of stable rules makes the modelling of learning of minor importance. Given the computing requirements of the AIE model, the scope of the thesis is limited to one relative stable period to avoid the overly onerous computing requirements to calibrate and predict over numerous periods and the additional requirement to model phase changes between the stable periods and their relation to changes in rules or learning. This rather major extension to the AIE model is left for further research, and will prove to be more practical in the future when the computing power has improved. Section 5.6.5, in further research, further discusses the phase change concept.
Second, AIE models a mature system, hence individual learning may have little relevance as section 2.2.2 discusses Ormerod et al.’s (2007, pp. 208-9) observation about how the more mature a system is, the more important the structure of the network is in determining the emergent behaviour than the intelligence of the individual agent. Consistently, Axtell and Epstein (1999, p. 177) note, ‘very little individual rationality may be needed for society as a whole ultimately to exhibit optimal behaviour.’ There is a large theoretical and experimental literature investigating how rational humans are, but from a network dynamic perspective this may be immaterial. This observation changes the focus from individual learning and raises the question ‘How does a network structure reflect learning?’ This is an interesting question, but it is beyond the scope of the thesis. The thesis makes the assumption that the profit expectations network is mature in between the structural breaks because the players have had chance to learn the rules.
Third, Figure 2–6 shows a persistent gap between the actual profits and expected profits, reflecting an optimism bias. This gap suggests that agents fail to accurately learn the true state of the world. The gap was one reason for abandoning the Bayesian learning approached, as discussed in section 3.1.