We first compare our work with other studies of the bullwhip effect with multi-agent systems. In fact, the nearest approach is fromKimbrough et al.[2002] who use a multi- agent system and a genetic algorithm to find a good ordering scheme for Sterman
[1989]’s Beer Game. It is close to our approach because it used multi-agent concepts with a model of company from the Beer Game. From this point of view, the only difference is in the structure of the supply chain, straight in the Beer Game versus divergent in the QWSG. However, there are many other differences. The main one is that they fix an ordering scheme and the genetic algorithm looks for the optimal value
of the parameter in this scheme. This optimal value is calculated so that the overall cost of the supply chain is minimized for the duration of the simulation. The optimum may therefore change depending on the simulation duration. On the other hand, our work focuses on the design of the ordering scheme itself, and we do fewer efforts to find optimal parameters, because we rely on the Solver in Excel fromMicrosoft Corp.[2004a], while they implement an optimization algorithm (genetic algorithm) by themselves. Conversely to their work, our ordering schemes are not designed to be efficient, but to stabilize ordering flows and inventories. We focus on efficiency only after this, when we optimize parameters in our two ordering schemes, but this is the essence of our approach. Though, these parameters are determined like Kimbrough and his colleagues, that is, they are set up such as the overall supply chain cost is a minimum for a specific duration of the simulation (but not for any duration).
The second approach using multi-agent systems is from Yung and Yang [1999a,b] who, like us, seek to reduce the bullwhip effect by improving the visibility and taking constraints into account. Instead of basing their model on the Beer Game, they use a model very close to the reality that can be operated by managers in an actual supply chain. Conversely to our higher level approach, they model processes in companies, but the supply chain considered in their experiments has only two levels (manufacturers and warehouses, where customers buy from these warehouses), while we consider four levels (retailers, wholesalers, paper and sawmill, and forest). Conversely to our agents applying an ordering scheme, their agents are more complex because they are able to:
• process routine jobs in place of humans on real installations;
• communicate information about operations, and use this information to optimize decisions with genetic algorithms;
• give managers a view of the state of the whole supply chain state, and in particular, a detection of the bullwhip effect, and an extraction of useful information received from other companies by a technique of constraint propagation.
The last approach with multi-agent systems is Yan [2001]’s study of the impact of delay distribution on the bullwhip effect. As it focuses on impact of delays in the supply chain, this work is very similar to ours. Basically, this agent-based supply chain checks experimentally that Chen et al. [2000]’s formal quantification of the bullwhip effect is correct. In particular, they verify like us that information centralization improves the efficiency of the overall supply chain.
We now compare our work with research studying the impact of delays on the bullwhip effect. All the following work focuses on the difficulty of forecasting future
demand when delays are longer, because the forecast has to be made on a longer horizon, conversely to our view, in which the bullwhip effect is also increased when there are no forecasts, because delays are a cause of the bullwhip effect by themselves. In this context, the approach from Chen et al. [2000], mentioned above, proposes a formal model in which the bullwhip effect is both induced by a forecasting technique, a moving average, and aggravated by delays. This paper was extended for another forecasting technique, an exponential smoothing, by Ryan [1997] and Simchi-Levi et al. [2000].
Lee et al. [1997a,b] also explain how forecasting techniques induce the bullwhip effect,
and how this is aggravated when delays are longer.
In general, such models are based on the Beer Game. As we use a simulation model derived from this game, we do so, but in a more empirical way. The Beer Game was studied by Sterman [1989] with human players, while we replace people by software agents. Therefore, we do not focus on players’ psychology that cannot understand the whole dynamics of the supply chain. Nevertheless, we also have to take such dynamics into account when proposing the ordering schemes used by our agents. As presented in this thesis, the design of our ordering schemes addresses the issue of delays in supply chains, because it is the only chain dynamics explaining the apparition of the bullwhip effect in the QWSG and the Beer Game played by software agents. From this point of view, we give a better understanding of Sterman’s “supply chain dynamics”, so that players should know how to play the Beer Game more efficiently. On the other hand, issues related to delays are very well known in supply chain management. In particular,
Sjöström[2001] studied their impact in the North European forest industry. This work
is similar to ours because it addresses a forest industry, but ours is less “empirical” as it uses simulation rather than real-life.
The last point of comparison that we consider is information sharing in supply chains, which has been well studied [Anderson and Morrice,2000;Cachon and Lariviere,
1999; Chatfield, 2001; D’Amours et al., 1999; Lee and Whang, 1998; Yu et al., 2001].
In fact, we propose to share information to reduce the bullwhip effect induced by delays, but information may be shared for other reasons. In our proposition of a mechanism for sharing information, we were inspired byPorteus [2000]’s Responsability Tokens to design our solution, even if our ordering scheme is eventually very different from this mechanism. In fact, Responsability Tokens deal with transfer payments in the supply chain to give incentives to upstream suppliers to ship products required downstream by retailers. On the contrary, our ordering scheme does not rely on money, because its only objective is to give enough information to companies, so that they can stabilize orders without neglecting operational constraints, and in particular inventory manage- ment. In fact, information sharing is the most often proposed solution to the bullwhip effect [Chen et al., 2000; Lee et al., 1997a; Simchi-Levi et al., 2000]. In this context,
our two principles improve the understanding of why and how to share information to reduce the bullwhip effect.
Another type of information sharing, supplier capacity information, is studied by
Swaminathan et al. [1997]. But from a more general point of view, two subfields of
Economics can be applied to study information sharing. The first subfield is Informa- tion Economics which studies the “consequences for the character and the efficiency of the interaction between individuals or organizations when one party has more or better information on some aspect of the relationship”. This is the condition of asym- metric information, under which the information gap will be exploited if, by doing so, the better-informed party can achieve some advantage” [MacHo-Stadler et al., 2001]. In particular, Cachon and Lariviere [1999] studied the value of information in supply chains. To illustrate information economics, let us assume that using our schemes was not Nash equilibria for the supply chain. In such an event, we could have given a price to the market consumption information. The question would have been, what minimal price would be demanded by retailers to accept to broadcast the market consumption information, and what maximal price other companies would agree to pay for this in- formation. Depending on these minimal and maximal prices, we would have been (or not) able to fix a price to the market consumption information, such as our schemes become Nash equilibria.
The second economics subfield that can be applied to supply chain management is game theory used to study incentives [Umbhauer,2002]. In particular,Cachon and Netessine
[2003] gave an overview of such studies and notice the “recent explosion of game- theoretic papers in supply chain management”. Our work belongs to this type of work, except that we replace the analytical model by a multi-agent model that we simulate, as suggested by papers in the multi-agent system field [Boutilier et al., 1994;
Rosenschein and Zlotkin, 1994; Sandholm,1999]. This approach gives us the ability to
solve more complex sets of relations.