یادگیری سازمانی در شبکه های تولید

If one accepts that a firm’s behavior is determined by history-dependent capabilities that adapt in a goal-directed way, one would like to know how a firm’s organizational structure influences the manner in which this distributed and partially tacit organizational memory evolves over time. In this paper, we study the impact that alternative information systems, incentive systems and modes of learning coordination have on the efficiency and generality of priority rules for job-shop scheduling which are learnt by a network of production agents modeled by neural networks. When modeling the alternative organizational structures by different input layers, feedback and training methods, we find that efficient rules evolve when global incentives and synchronized learning are employed even if the system state is only partially known to an agent. However, organizational learning fails when it is performed asynchronously with local goals.

Traditionally, economics judges an economic organization according to the efficiency of allocation obtained when employing it. There, an agent has to solve a given problem of optimization under constraints derived from the agent’s capabilities and actions of other agents. Coordination of the actors is achieved either via the market mechanism through the adjustment of supply and demand or through the firm employing an organizational structure that provides problem decomposition, communication channels, and incentive mechanisms. The choice between coordination mechanisms is made according to the transaction costs involved, where the production costs are assumed to be unaffected by organizational arrangements (see, e.g. Williamson, 1975). A number of works on the different aspects of a firm’s organization using this paradigm indicate that the hierarchy is the best way of keeping a firm’s transaction costs low, as it economizes the costs of information and communication (Bolton and Dewatripont, 1994) and allows optimal coordination (Mesarovic et al., 1970). Consequently, Williamson titles his book on transaction cost economics “Markets and Hierarchies” equating the “firm” with “hierarchy”. They describe a firm as a hierarchically ordered set of automata that process information in order to derive organizational actions. In the long run, only firms with the most efficient production and governance mode survive under perfect competition, i.e. differences between firms are transient and profits vanish except in the presence of market power. Hence, the purpose of strategy is the selection of suitable markets, the necessary organizational set-up being viewed as a subsequent step (Porter, 1985). When looking at the software systems used in practice for production planning and control, one finds that this “information processing paradigm” is a rather good description of how firms operate: based on given product and process specifications laid down in bills of materials and work plans, enterprise resource planning (ERP) software packages typically start with coordinating sales and logistics via aggregate demand planning; on this basis, they then coordinate the production segments among each other and with the purchasing function by deriving production orders and purchase requirements which are subsequently detailed into schedules and orders, respectively (see, e.g. Scheer, 1998). However, the “information processing paradigm” neglects the aspect of innovation and thus cannot explain a number of empirical observations such as follows. • Lawrence and Lorsch find that real-world organizations are the less hierarchical the more often they have to innovate. They conjecture that the hierarchy is not always the best way of organizing. Rather, this is contingent on the environment (Lawrence and Lorsch, 1986). • Nelson reports persisting and substantive differences in organizational forms and performance among firms rather than a convergence to a single form (Nelson, 1991). • Teece et al. show that successfully diversifying firms only move in areas close to the current field of operations, suggesting that a firm’s strategy is historically constrained by the capabilities obtained in the current area of operations (Teece et al., 1994, see also Pennings et al., 1994). Explanations for these phenomena are offered by evolutionary approaches. They view the firm as a “repository of capabilities” for developing, producing, and selling products which determine the current and possible future behavior (Loasby, 1998). Rather than static or publicly available, capabilities are largely tacit and have to be acquired in an idiosyncratic and path-dependent way via social learning by doing and imitation. They can be seen as the distributed “organizational memory”, consisting of declarative frames of reference that determine how the organization “sees the world” and of procedures for performing tasks and coordination, both of which adapt incrementally to own or imitated foreign experience on the basis of feedback in a target-oriented way. From this point of view, a firm’s organizational structure is to be judged according to the long-term competitiveness of the organizational memory that evolves through it, i.e. depending on how well the organization learns. Not surprisingly, the suggestions for organizing a firm from this point of view differ substantially from those derived from the transaction cost point of view. When people are presented with an environment that has structure, tacit knowledge is created to exploit its structure to guide behavior (Reber, 1993). Thus, every organization member develops valuable insights through learning by doing. For the integration and explication of tacit knowledge figures, analogies and metaphors are used (Kim, 1993 and Nonaka and Takeuchi, 1995) and thus high-bandwidth communication channels are needed rather than the low bandwidth required for communicating prices, quantities, or due dates in the information processing paradigm (Draft and Lengel, 1986). The definition of new problems via the integration and explication of learnt declarative knowledge and its embodiment in new routines is hence best accomplished in multifunctional teams whose members are also from operations. For this purpose, Nonaka and Takeuchi propose a “hypertext organization”: exploitation of existing knowledge in operations is coordinated hierarchically, while coordinated explication is accomplished by teams (Nonaka and Takeuchi, 1995). Total quality management (TQM) offers guidelines for this process. The House of Quality for instance provides a graphical representation of mental models regarding functional relationships between and within customer preferences, technical specifications, and process designs that are explicated and enriched with quantitative data by a multifunctional development team and are, subsequently, used as a fitness landscape when searching for new routines (see, e.g. Winter, 1996 and Hauser and Clausing, 1988). Similarly, the “seven tools for Kaizen” are graphical methods for coordinating the continuous detection of cause–effect linkages and subsequent application of this knowledge for process improvement (Lillrath and Kano, 1989). Management guides this process through “vision” and “cognitive leadership” (Witt, 1998). The vision gives the intention via describing what knowledge should be generated. It has to be considerably more general and vague than a strategy derived from game-theoretic arguments, as it has to cope with the fact that innovations are unpredictable and to allow the integration of the more detailed organizational knowledge generated by the other organization members (Winter, 1987). Typical visions are “Telecommunication, Global, Focus” by Nokia, or “The Network is the Computer” by Sun Microsystems (Brown and Eisenhardt, 1998, p. 148). Cognitive leadership is the social competency of communicating the vision and sustaining it as the tacit cognitive framework collectively shared by the firm’s members. The role of incentive systems in a learning organization is to provide positive feedback to experimenting, reflecting, and knowledge sharing rather than to induce an agent whose actions are imperfectly observed to perform as close as possible to the optimum according to the principal’s optimization model. Promotion and/or premia based on achievements of design or improvement teams and profit-sharing schemes seem appropriate for this purpose. This is shown e.g. by Minkler (1993) who demonstrates that a profit-sharing scheme arises when the agent has more knowledge than the principal even in the case of perfect information. Also, the proper framing via the vision and the social role of leaders contribute to tackling opportunism. Nonaka and Takeuchi view autonomy, fluctuation, and creative chaos as well as redundancy of information as enablers of organizational learning—all concepts that increase transaction costs when compared to a purely hierarchical structure (see Nonaka and Takeuchi, 1995).

We investigated various aspects of a hypertext organization for job-shop scheduling in a production network in the sense that scheduling routines evolve under coordinated learning in an exploration phase and are then routinely applied in an exploitation phase. The main findings of our inquiry into this organizational learning problem can be summarized as follows. • Uncoordinated learning with local goals is inferior as, under this setting, it is hard to find proper cause–effect relations. It seems appropriate only for obtaining a fast initial improvement. If this organizational structure is chosen, one should find a learning rate that is a good compromise between ambiguous learning and learning too slowly. In general, a lower learning rate is advantageous. In this respect, our findings coincide with those by Lounamaa and March (1987). • Coordinated learning has its price in the form of a higher coordination effort. For the problems studied in this paper, it turned out that the strongest coordination—namely the learning of one commonly applied rule based on synchronized experimentation and global feedback—yielded the best results. Improving this setting locally did not improve the results. Thus, in this case the diversity in the distributed organizational memory was not advantageous. This finding can be one reason for the recently emerged interest in supply chain management which stresses the need to centralize the planning and control of production and logistics for firms operating in a network of supply relations (see, e.g. Poirier, 1999). • In most cases, it was found that a strategy with foresight was beneficial. Taking into account also the previous results, we suggest that incentives based on local machine utilization are not good for organizational learning. Altogether, the results obtained strongly support the notion that a process-oriented organization coupled with teams that coordinate and evaluate continuous improvement as described, e.g. in Monden (1995) or Lillrath and Kano (1989) is the best way of organizing organizational learning. • On the other hand, the information system used can be quite parsimonious. In general, good results are obtained by a global static information system where the information about the operations is generated when the load is passed on to the production network. Furthermore, the information can be moved together with the parts to be produced via the operations to be scheduled, and on this basis, the decisions in the exploitation phase can be made autonomously. Also, the additional effort expended on the communication needed by a strategy with foresight only requires the local exchange of data between two agents. • The results obtained contradict the analytical separation between production and governance structure made in transaction cost economics. In our case, production costs are a function of the quality of the scheduling rules learnt which depends on the mode of governance used (see Hodgson, 1998 for a similar line of reasoning). Clearly, while having empirical support and correlating to other similar works, the results described above only hold for the problem studied in this paper. Thus, the development of analytical models of organizational learning would be a fruitful field of further research. Other areas of interest could be the explicit modeling of the coordination needed for integrating the distributed organizational memory into a common one when a single rule shall be evolved as well as the study of combined autonomous-coordinated learning schemes.