Complex Decision-Making in Economy and Finance. Pierre Massotte
interactions that may occur at long distances.
Based on these remarks, the following strategies can be formulated.
1.4.2. Interactions in industrial workshops
On the functional and logistical level, an industrial system has a network structure, as found in ecosystems or biology, with medium-sized non-hierarchical aggregations. At the information system level, this system contains positive and negative feedback loops: this is the case when several activity centers (workshops) are interconnected to form a complete plant. Such a system, with its interaction and feedback loops, is presented in Figure 1.1.
Figure 1.1. An industrial MRP system with its feedback loops
More formally, it is a cybernetic system with strong interactions between distributed functions: a given function will influence the activity or “inhibition” of a neighboring function. The problem involves analyzing the overall behavior resulting from such a system; a systems dynamics approach is appropriate. We have applied it to study the dynamic influence of interactions between various manufacturing plants of the type: Semiconductor – Electronic boards – Computers.
This set is subject to stepwise requests, at the final product “computers” level. The objective here is to analyze the impact of disruptions (demand level, yield variations, “over-reaction of production agents”) on the evolution of inventories and work in progress upstream, at the card and then semiconductor level. Modeling in the form of differential equations was carried out in a simple way on such systems, and it was possible to easily integrate the parameters and variables (size of buffer stocks, response times, etc.). In this case, it has been shown [MAS 95c] that the variation in component stocks, i.e. upstream, becomes chaotic when the system is subjected to simple control variations. As the capacity of the production system is limited, some products are penalized compared to others. This situation occurs more precisely when the flow of information is amplified and when we try to “recover” the drifts because the production agents tend to anticipate disruptions. On the contrary, in the case where the input variable is chaotic, it has not been possible to demonstrate with certainty that the system was chaotic: because of the buffer stocks (number, size, response time, etc.), the results can be smoothed, attenuated or even amplified. When the size of the buffer stocks is large and the yields are low, it has been found that the system can even diverge: in this case, the noise is amplified and directly influences the evolution of the system. However, this is also not possible due to the lack of sufficient reliable data.
Conventional management systems are essentially “planning”: they are intended to prepare a production system but not to manage it as well as possible. In some cases, they are even useless because they are too restrictive, sometimes unable to control a complete production system. In addition, they are subject to the normal reaction of a planning agent who reacts in the opposite direction to certain trends in an attempt to compensate for variations. In this case, the overly “planning” and “rigid” management system leads to dangerous, unexpected and unpredictable variations, such as changes and jolts that are amplifying disturbances. On another level, conventional production management systems are not designed to handle chaotic programs. In such production systems, reactive management tools can only degrade their overall performance. We therefore deduce that planning management systems, which are so reassuring and very useful in other respects, are costly in terms of procedures (many steering meetings, production readjustments, multiple planning steps, etc.) and are not the best suited to certain types of dynamic behavior.
According to the well-known principles of stability, it is sufficient to introduce compensation loops or decision-making centers whose action will change the direction of variation, positive or negative, of the product or information flow. It is also possible to control unstable systems by injecting “noise” and uncertainty into control parameters as well as stimuli and input variables. This allows the possibility of compensating or even eliminating knocks and counteracting the effects of pumping. For example, it is common for decision-makers to introduce noise: they can constantly modify product priorities in the workshop, depending on changes in the situation and conflicting demands. In fact, the stimuli generated are based on their perception of the problem and tend to create more disruptions and blows.
1.4.3. Product flow in a flexible production system
Consider a flexible multi-product and multi-process workshop with duplicated equipment, feedback loops and complicated ranges.
Figure 1.2. A flexible production system
Each node or cell has its own control system and behavioral procedures. We are in a “local” environment with a limited proximity; the production rules used here concern priority management, order sequencing, alarm management (in economy and finance: alert management) in case of problems, etc.
Such a production system has been studied [BAR 96] in terms of the organization to be put in place to compensate for the disruptive effects of chaos (in the sense that they are unpredictable). Here, the chaos is essentially due to interactions between cells: oscillations created by calls, supply orders that propagate from cell to cell, going up the “line” of manufacture. They are also production or launch orders from the production management system, which will spread from one cell to another and vice versa. This will generally be done from downstream to upstream, if we operate in a pull flow. When these orders respond to nonlinear functions or influences and the phenomena are amplified, making the system sensitive to initial conditions (SIC), this induces many possible states for the production system, which can be the result of deterministic chaos. This is observed, for example, in the dynamic variation of stocks (WIP or Work In Process) throughout the “line”. The cause is due to the sequencing and amplification effects specific to the physical and logical structure of the production system. We will call this the caterpillar effect. Under these conditions, we cannot predict the behavior of such a system. Moreover, the model corresponding to a real workshop is relatively complex and cannot integrate all the parameters and assumptions: it cannot be used for steering purposes. Simulation can therefore be used to “gauge” a complex system, to evaluate trends and define the least bad strategies.
In terms of the strategies adopted, one of them consists of returning the production system to a stable state, which amounts to placing it in an area of the known and “reassuring” phase space corresponding to an area of weak bifurcations. However, its adaptability is reduced because, to reconfigure it and bring it into a new given state, it will require a lot of energy (since its inertia is greater). On the contrary, we will recommend exploiting the chaotic nature of such a system. More generally, we must seek conditions that place the production system at the limit of stable and unstable states. In an area of “weak chaos”, which is easy to achieve in nonlinear workshops, its flexibility is maximum. These concepts formed the basis of an international project called GNOSIS-VF (EU ESPRIT 28448 project) as part of the Intelligent Manufacturing Systems (IMS) program with Japan.
In this example, behavioral complexity is related to the presence and importance of interactions between the different agents that make up the production system. Here, simple deterministic functions applied to strongly linked systems can generate chaos. In this case, it is possible to control its effects by decoupling the cells from the system through a double Kanban system; this way, the value of the work in progress can be limited while leaving each cell its own elasticity and having disturbances that compensate themselves. With reduced buffer stocks, the adaptation of inputs and outputs is rather rapid thanks to self-regulatory effects; the best strategy is then to let the system evolve freely, maintaining the parameters