SKIN (Simulating Knowledge Dynamics in Innovation Networks) model
- 1 SKIN – Simulating Knowledge Dynamics in Innovation Networks
- 1.1 Overview
- 1.2 Design concepts
- 1.3 Details
- 1.4 Implementation
- 1.5 Bibliography
- 1.6 The code
SKIN – Simulating Knowledge Dynamics in Innovation Networks
This model description follows the ODD protocol for describing individual- and agent-based models (Grimm et al 2006) As required by the protocol, the description provides an overview, the general concepts of the model, and some details. A final section (not part of ODD) describes how to run the model
The SKIN (Simulating Knowledge Dynamics in Innovation Networks) model is an agent-based simulation representing a theory of the dynamic processes involved in innovation in modern knowledge-based industries (Gilbert et al 2001; Pyka et al 2002; Ahrweiler et al 2004; Ahrweiler et al 2006; Pyka et al 2007; Gilbert et al 2007). The agent-based approach allows the representation of heterogeneous agents that have individual and varying stocks of knowledge. The simulation is able to model uncertainty, historical change, effect of failure on the agent population, and agent learning from experience, from individual research and from partners and collaborators. The aim of the simulation is to show that the artificial innovation networks share certain characteristics with innovation networks in knowledge intensive industries that are difficult to accommodate in traditional models of industrial economics.
State variables and scales
SKIN is a multi-agent model containing heterogeneous agents which act and interact in a complex and changing environment. The agents represent innovative firms who try to sell their innovations to other agents and end users but who also have to buy raw materials or more sophisticated inputs from other agents (or material suppliers) in order to produce their outputs. This basic model of a market is extended with a representation of the knowledge dynamics in and between the firms. Each firm tries to improve its innovation performance and its sales by improving its knowledge base through adaptation to user needs, incremental or radical learning, and co-operation and networking with other agents. The model is intended to represent one technological knowledge-intensive industrial sector (such as the biotechnology sector). Although for simplicity the agents are referred to as ‘firms’ in the following description, agents can represent all organisational actors, including for example research institutes, university research groups, and management consultancies. Similarly, although the knowledge held and exchanged by the firms will typically be technical knowledge, the model covers all kinds of knowledge, including, for instance, knowledge about marketing methods, packaging techniques, and even legal and regulatory knowledge.
Process overview and scheduling
In the model, a firm has several ways of improving its performance, either alone or in co-operation, and in either an incremental or a more radical fashion. All strategies have in common that they are costly: the firm has to pay a “tax” as the cost of applying an improvement strategy. The firms continue to perform until their capital is exhausted, at which time they ‘die’ and are removed from the model. If the sector is successful, existing firms may enter the sector or new ones may be created: such firms are called ‘start-ups’.
The core concept of the framework is the knowledge which will manifest itself in the innovative production or delivery of manufactured and service products. The approach to knowledge representation used in the model is similar to Toulmin’s (1967) evolutionary model of knowledge production. This identified concepts, beliefs and interpretations as the "genes" of scientific/technological development evolving over time in processes of selection, variation and retention. In the SKIN model, a “kene” is used to represent the aggregate knowledge of an organization (Gilbert 1997).
The individual knowledge base of a SKIN agent, its kene, contains a number of “units of knowledge”. Each unit is represented as a triple consisting of a firm’s capability C in a scientific, technological or business domain (e.g. biochemistry), represented by an integer, its ability A to perform a certain application in this field (e.g. a synthesis procedure or filtering technique in the field of biochemistry), represented by a real number, and the expertise level E the firm has achieved with respect to this ability (represented by an integer). The firm's kene is its collection of C/A/E-triples which is of variable size and represents an artificial knowledge space.
Figure 1 The kene of a firm
When it is set up, each firm has also a stock of initial capital. It needs this capital to produce for the market and to improve its knowledge base, and it can increase its capital by selling products. The amount of capital owned by a firm is a measure of its size and also influences the amount of knowledge that it can support, represented by the number of triples in its kene. Most firms are initially given a starting capital allocation, but in order to model differences in firm size, a few randomly chosen firms can be given extra capital. In many knowledge-intensive industries we find a co-existence between large and small actors (e.g. the large pharmaceutical firms and the biotech start-ups, or the former national monopolists and high technology specialists in the ICT-industries). This particular distribution of firm sizes makes it necessary to discriminate between large and small actors in the simulation set-up.
Firms apply their knowledge to create innovative products that have a chance of being successful in the market. The special focus of a firm, its potential innovation, is called an innovation hypothesis. In the model, the innovation hypothesis (IH) is derived from a subset of the firm’s kene triples.
Figure 2 Forming an innovation hypothesis
The underlying idea for an innovation, modelled by the innovation hypothesis, is the source an agent uses for its attempts to make profits on the market. Transforming the innovation hypothesis into a product is a mapping procedure where the capabilities of the innovation hypothesis are used to compute an index number that represents the product. The particular transformation procedure applied allows similar products resulting from different kenes, which is not too far from reality where the production technologies of firms in a single industry vary considerably. A firm’s product, P, is generated from its innovation hypothesis as
(where N is a constant). The product has a certain quality which is also computed from the innovation hypothesis in a similar way, by multiplying the abilities and the expertise levels for each triple in the innovation hypothesis and normalising the result. In order to realise the product, the agent needs some materials. These can either come from outside the sector (“raw materials”) or from other firms, which generated them as their products. What exactly an agent needs is also determined by the underlying innovation hypothesis: the kind of material required for an input is obtained by selecting subsets from the innovation hypotheses and applying the standard mapping function (see equation 1 above). These inputs are chosen so that each is different and differs from the firm’s own product. In order to be able to engage in production, all the inputs need to be obtainable on the market, i.e. provided by other agents or available as raw materials. If the inputs are not available, the agent is not able to produce and has to give up this attempt to innovate. If there is more than one supplier for a certain input, the agent will choose the one at the cheapest price and, if there are several similar offers, the one with the highest quality.
Figure 3 A firm’s input requirements
Input 1: (A1 + A2) modulus N
Input 2: (A3 + A4 + A5) modulus N
If the agent can go into production, it has to find a price for its own product which takes account of the input prices it is paying and a possible profit margin. While the simulation starts with product prices set at random, as the simulation proceeds, a price adjustment mechanism increases the selling price if there is much demand, and reduces it (but no further than the total cost of production) if there are no customers. A range of products are considered to be ‘end-user’ products and are sold to customers outside the sector: there is always a demand for such end-user products provided that they are offered at or below a fixed end-user price. An agent will then buy the requested inputs from its suppliers using its capital to do so, produces its output and puts it on the market for others to purchase. Using the price adjustment mechanism, agents are able to adapt their prices to demand and in doing so learn by feedback.
In making a product, an agent applies the knowledge in its innovation hypothesis and this increases its expertise in this area. This is the way that learning by doing/using is modelled. The expertise levels of the triples in the innovation hypothesis are increased by 1 and the expertise levels of the other triples are decremented by 1. Unused triples in the kene eventually drop to an expertise level of 0 and are deleted from the kene; the corresponding abilities are “forgotten” or “dismissed” (cf. e.g. Hedberg 1981).
If a firm’s previous innovation has been successful, i.e. it has found buyers, the firm will continue selling the same product in the next round. However, if there were no sales, it considers that it is time for change (evaluating feedback). If the firm still has enough capital, it will carry out “incremental” research (R&D in the firm’s labs).
Performing incremental research (cf. Cohen and Levinthal 1989) means that a firm tries to improve its product by altering one of the abilities chosen from the triples in its innovation hypothesis, while sticking to its focal capabilities. The ability in each triple is considered to be a point in the respective capability’s action space. To move in the action space means to go up or down by an increment, thus allowing for two possible “research directions”.
Initially, the research direction of a firm is set at random. Later it learns to adjust to success or failure: if a move in the action space has been successful the firm will continue with the same research direction within the same triple; if it has been a failure, the firm will randomly select a different triple from the innovation hypothesis and try again with a random research direction.
Figure 4 Incremental research
A firm under serious pressure that is in danger of becoming bankrupt will turn to more radical measures, by exploring a completely different area of market opportunities. In the model, an agent under financial pressure turns to a new innovation hypothesis after first “inventing” a new capability for its kene. This is done by randomly replacing a capability in the kene with a new one and then generating a new innovation hypothesis.
An agent in the model may consider partnerships (alliances, joint ventures etc.) in order to exploit external knowledge sources. The decision whether and with whom to co-operate is based on mutual observations of the firms, which estimate the chances and requirements coming from competitors, possible and past partners, and clients.
The information a firm can gather about other agents is provided by a marketing feature: to advertise its product, a firm publishes the capabilities used in its innovation hypothesis. (Capabilities not included in its innovation hypothesis and thus in its product, are not visible externally and cannot be used to select the firm as a partner.) The firm’s advertisement is then the basis for decisions by other firms to form or reject co-operative arrangements.
In experimenting with the model, we can choose between two different partner search strategies, both of which compare the firm’s own capabilities as used in its innovation hypothesis and the possible partner’s capabilities as seen in its advertisement. Applying the conservative strategy, a firm will be attracted by a possible partner who has similar capabilities; using a progressive strategy the attraction is based on the difference between the capability sets.
Previously good experience with former contacts generally augurs well for renewing a partnership. This is mirrored in the model: to find a partner, the firm will look at previous partners first, then at its suppliers, customers and finally at all others. If there is a firm sufficiently attractive according to the chosen search strategy (i.e. with attractiveness above the ‘attractiveness threshold’), it will stop its search and offer a partnership. If the possible partner wishes to return the partnership offer, the partnership is set up.
The model assumes that partners learn only about the knowledge being actively used by the other agent. Thus, to learn from a partner, a firm will add the triples of the partner’s innovation hypothesis to its own. For capabilities that are new to it, the expertise levels of the triples taken from the partner are reduced by 1 in order to mirror the difficulty of integrating external knowledge (cf. Cohen and Levinthal 1989). For partner’s capabilities that are already known to it, if the partner has a higher expertise level, the firm will drop its own triple in favour of the partner’s one; if the expertise level of a similar triple is lower, the firm will stick to its own version. Once the knowledge transfer has been completed, each firm continues to produce its own product, possibly with greater expertise as a result of acquiring skills from its partner.
If the firm’s last innovation was successful, i.e. the amount of its profit in the previous round was above a threshold, and the firm has some partners at hand, it can initiate the formation of a network. This can increase its profits because the network will try to create innovations as an autonomous agent in addition to those created by its members and will distribute any rewards to its members who, in the meantime, can continue with their own attempts, thus providing a double chance for profits.
Networks are “normal” agents, i.e. they get the same amount of initial capital as other firms and can engage in all the activities available to other firms. The kene of a network is the union of the triples from the innovation hypotheses of all its participants. If a network is successful it will distribute any earnings above the amount of the initial capital to its members; if it fails and becomes bankrupt, it will be dissolved.
If a sector is successful, new firms will be attracted into it. This is modelled by adding a new firm to the population when any existing firm makes a substantial profit. The new firm is a clone of the successful firm, but with its kene triples restricted to those in the successful firm’s advertisement, and an expertise level of 1. This models a new firm copying the characteristics of those seen to be successful in the market. As with all firms, the kene may also be restricted because the initial capital of a start-up is limited and may not be sufficient to support the copying of the whole of the successful firm’s innovation hypothesis.
All firms are initialised with a random kene (i.e. a set of triples with values chosen from uniform random distributions) and with initial capital set as a parameter (see below). Consequently, in the first few time steps (up to about time step 50, but depending on the level of initial capital) the model demonstrates a severe shakedown, as those firms that are unable to make a saleable product and thus obtain an income steadily lose their capital and eventually become bankrupt and die. This shakedown period is unrealistic (the only example of a sector having many firms that are unable to sell their products is on the occasion of a major exogenous disruption, such as the 1970s oil price rise, or the imposition of a new draconian regulation), and so normally observations of the model are only started following the shakedown period.
The standard scenario uses the following parameter settings:
- Initial capital: 20000 (large firms: 200000)
- Initial population of firms: 500
- Number of large firms, with extra capital at the start: 50
- Range of product index numbers in the sector: ]0.0 to 100.0[
- Maximum difference between product and input index numbers for them to be considered substitutable: 1.0
- All products with a product number below 5.0 are considered to be ‘raw-materials’ and all those with numbers above 95 are ‘end-user’ products.
- Price of raw materials: 1
- Maximum price of end-user products: 1000
- Profit required attracting new start-ups: 1200
- Partnering search strategy: conservative
- Attractiveness threshold to allow two firms to partner: 0.3
- Capital cut-off below which firms do radical rather than incremental research: 1000
- Taxes: per time step: 200; per incremental research attempt: 100; per radical research attempt: 100: per collaboration partner: 100
There are no submodels.
The model has been programmed using NetLogo (http://ccl.northwestern.edu/netlogo/), version 3.14.
Figure 5: The interface of the model
The main graphic window (Figure 5) shows about 100 firms (represented by the small ‘factory’ shapes). Their position in the display window is not significant: a layout algorithm is used to move the factory icons to positions where they can best be seen. Factory icons on their side mark firms that were created after the start of the simulation (i.e. they are start ups) and those that are upside down are network firms (firms producing on behalf of a network, with a kene based on the union of the kenes of the network members). The size of the icons indicates the amount of capital its firm possesses (the size is proportional to log10 of the amount of capital).
The numerous lines indicate partnerships, supplier relationships and network linkages between firms. Those not involved in any relationship have been moved by the layout algorithm to the margins of the display. There are 5 networks (one with 4 members and four with 3 members). The networks are shown with lines interconnecting all their members. The medium grey factories are those producing ‘raw materials’ (these firms require no inputs from within the system in order to create their products); the dark grey ones are ‘end-users’ that produce no outputs.
The display is surrounded by graphs monitoring various aggregate aspects of the system. At the top right is shown the growth in the population of firms as start-ups are added, and the slow growth in the number of networks. The graph below shows the percentage of firms that have products on the market (‘Firms selling’) and the percentage that have made a sale (‘Sales’); the latter is always less than the former because some firms are unable to find customers prepared to buy at the price proposed for the product. The third graph down the right-hand side indicates the percentage of firms that are involved in either at least one partnership or in a network. The bottom right graph shows a measure of the distribution of funds in the market, the Herfindahl concentration index Ht,
where sti is the relative capital of firm i, which measures the distribution of capital among the firms.
The Networks histogram shows the number of firms in each network, and the Dynamics plot indicates, on the upper Successes graph, the number of firms which have exceeded the threshold of profit that indicates a successful innovation (the ‘success threshold’) and, on the lower Start Ups graph, the number of new firms entering the market at each round. The Capital plot in the bottom left corner shows the average capital of the firms, expressed as a logarithm base 10.
To run the model, set the sliders to the desired parameter values, and press the Setup button, which initialises the model. Then press the Go button to start the simulation.
Ahrweiler, Petra, Gilbert, Nigel, & Pyka, Andreas. (2006) 'Institutions Matter But... Organisational Alignment in Knowledge-Based Industries'. Science, technology and Innovation Studies, 1(2), 39-58 http://www.sti-studies.de/articles/2006-01/ahrweiler.htm.
Ahrweiler, Petra, Pyka, Andreas, & Gilbert, Nigel. (2004) 'Simulating Knowledge Dynamics in Innovation Networks', In Leombruni, R. & Richiardi, M. (Eds.), Industry and Labor Dynamics: The Agent-Based Computational Economics Approach (pp. 284-296). Singapore: World Scientific Press.
Cohen, W.M. and Levinthal, D. (1989) ‘Innovation and learning: the two faces of R&D’, The Economic Journal, Vol. 99, pp.569–596.
Gilbert, N. G. (1997), ‘A Simulation of the Structure of Academic Science’. In: Sociological Research Online 2/2, http://www.socresonline.org.uk/socresonline/2/2/3.html.
Gilbert, Nigel, Ahrweiler, Petra, & Pyka, Andreas. (2007) 'Learning in Innovation Networks: Some Simulation Experiments'. Physica A, 378, 100-109 http://dx.doi.org/10.1016/j.physa.2006.11.050.
Gilbert, Nigel, Pyka, Andreas, & Ahrweiler, Petra. (2001) 'Innovation Networks - a Simulation Approach'. Journal of Artificial Societies and Social Simulation, 4(3)8, http://www.soc.surrey.ac.uk/JASSS/4/3/8.html.
Grimm V. et al (2006) A standard protocol for describing individual-based and agent-based models. Ecological Modelling 198, 115-126.
Hedberg, B. (1981): How Organizations learn and unlearn. In: P. C. Nystrom and W. H. Starbuck (eds.): Handbook of Organizational Design. Oxford: Oxford University Press, pp. 3-27.
Pyka, Andreas, Gilbert, Nigel, & Ahrweiler, Petra. (2002) 'Simulating Innovation Networks', In Pyka, A. & Küppers, G. (Eds.), Innovation Networks: Theory and Practice. Cheltenham: Edward Elgar.
Pyka, Andreas, Gilbert, Nigel, & Ahrweiler, Petra. (2007) 'Simulating Knowledge-Generation and Distribution Processes in Innovation Collaborations and Networks'. Cybernetics and Systems, 38(7), 667-693.
Toulmin, S. (1967): The Philosophy of Science. An Introduction. London: Hutchinson.