© Copyright JASSS
Bruce Edmonds (2001)
Commentary on: Sophie Thoyer, Sylvie Morardet, Patrick Rio,
Leo Simon, Rachel Goodhue and Gordon Rausser (2001)
A Bargaining model to simulate negotiations between water users
Journal of Artificial Societies and Social Simulation
vol. 4, no. 2,
To cite articles published in the Journal of Artificial Societies and Social Simulation, please reference the above information and include paragraph numbers if necessary
- What are the goals of this paper? The authors state that"...the predictive power of the model is not our main concern", rather they wanted to "...provide a better understanding of the complex interrelations between the various components of the modelled system". The purpose of this was broadly to improve negotiations, for example to"...provide guidelines to the negotiation organisers, helping them to choose the adequate negotiation structure", or to "...increase the participation of stakeholders..". I will argue that it fails in these goals - fails because the model structure is essentially unjustified (except by tradition). This breaks any explanatory chain and makes prediction on unseen data extremely unlikely.
- The core of its model is a sequence of games. The model is solved backwards from the last game to the first. A solution to the model is forced by stipulating that if there is not an agreement, the outcome is worse for the participants than any other outcome. The moves by the players are determined by concave utility functions whose forms are largely taken from the micro-economic tradition, but some of whose parameters are estimated (with some input from stakeholders).
- Thus although some of the parameters of the complete model originate from the domain of study, the structure is arrived at largely by assumption. The outcomes of the model are due either to the values of these parameters or the assumptions used in the construction of the model (or some mixture of both). The assumption throughout the paper is that the outcomes of the model are explained by the initial parameterisation, for example, a greater allocation of water to farmers was accounted for by their increased negotiation weight. In other words the authors have assumed that the structure of the model is essentially correct, that it is capable of faithfully transmitting the causes to the outcomes in a similar way to the real negotiations. The big question to be asked of this paper is whether there are grounds for supposing this.
The model is explicitly based upon many assumptions known to be false. In the model the negotiators judge whether to accept a round of negotiation by summing over all the (other) players' future proposals - proposals that they don't yet know. The negotiators make their decisions by recursively solving the negotiation by backwards induction starting from the last round and working their way to the first. In order to ensure that there is a solution to this a number of other (false) assumptions are forced upon the authors - thus the choice of the type of core model has the effect of further distorting the representation. I will briefly look at two such sets of assumptions.
- In the model, if there is no agreement in the last round a proposal that is worse for all the participants is forced upon them - this ensures that participants will accept a compromise however inimical to their interests. This would not be the case in reality (where the government steps in and dictates a compromise in absence of agreement), because negotiations do sometimes fail precisely because that will sometimes be to the advantage of particular participants (among other reasons).
- It is assumed that the policy alternatives form a convex subset and the players alternatives are strictly concave. Taken together this would mean that the negotiation process would be almost trivial. No novel proposals are possible, everybody knows in advance all possible negotiation moves, making alliances is pointless, and the whole process is reduced to a series of price suggestions for each variable - a sort of multidimensional haggling. The complexity of negotiations, the very complexity that makes them necessary, has been eliminated.
- In justification of their choice of model structure the authors say that"...the complexity of agents interactions has led us to adopt necessary simplifications which increase the gap between simulated game solutions and real world outcomes". While, inevitably, some simplifications are necessary for any model, this does not force one to choose inappropriate simplifications. The key choice here, namely the one to base this study on a series of artificial games, is contrary to what we observe in real negotiations and, ultimately, prevents the authors from attaining their goals.
- Thus we have to completely rely on the "validation" to verify the correctness of the model. This is done with some experiments to assess whether the model acts in an "expected" way when parameters are changed - a sort of basic sensitivity analysis. Thus the following is ascertained: that the model can give outcomes that are plausibly of the right order of magnitude; that negotiation outcomes are shifted to be more desirable for a participant as that participants "weight" is increased in a smooth and continuous fashion; that excluding or limiting the impact of a participant biases the final outcome away from that participants interests; and that the initial position influences the outcome. These simple tests give us some confidence that there are no serious bugs in the implementation (compared to its design) but (as the authors recognise) they are insufficient to give confidence that the model is a good representation of the real negotiation. A"second phase of the validation" would be necessary for this (as the authors recognise).
What new insights does the model give us?
- Given that, at the moment, there is no reason to suppose that this model reflects the reality of real negotiations, let us consider whether it can nonetheless provide us with some useful or new insights into a possible negotiation process. The point of the experiments reported in the paper were to verify the structure of the model. That the outcomes of these experiments were commensurate with what was expected, which was (correctly) taken as a weak confirmation that the model construction was sound. There is no emergence here - in other words, there is no indication that there is any behaviour that was not explicitly designed into it in the first place. This is not greatly surprising since the model they used was one designed to be a fairly simple one, since the authors were not seeking abstract results about the working of such games but rather seeking to apply it to negotiations about water.
Could it ever be the basis of a predictive model?
- This model is unlikely to predict unseen outcomes. This was not the authors' intention in building the model. However since they do not completely dismiss the possibility that it may be predictive let me point out that the model does make predictions that do not generally hold. These include: that the negotiations will always come to an agreement; that there will not be any sharp "hysteresis" in the outcomes w.r.t. the set-up (due to the assumptions discussed in (1) above); and that the presence of negotiators who do not participate forces other participants to "take their reaction into account" even though they have null weight.
Can this model supply us with any interesting or useful explanations of the observed phenomena?
- Given that the model does not take its structure from observations of the target and that it does not predict, how will it provide us with any explanations? The authors state several times that it is understanding of negotiations that the model is to help with, but they do not say how this understanding is to come about. Is an outcome to be explained in terms of a mechanism know not to exist in real negotiations? The trouble is that attempt to construct and explanation using the model are frustrated by the nature of the model. If a process that caused a certain outcome is traced back through the model machinery this becomes a prototype for an explanation, but it is only a viable explanation if it has a credible interpretation back into the real process. In this case it would be very difficult to interpret the simplistic mechanics of the backwards sequence of games onto the complex forward dynamics of a real negotiation. For this reason the only sort of explanation derivable from this model is one that assumes that the model reflects the real process somehow - an assumption that, in my opinion, is hard to justify.
Is it likely to be a helpful tool for aiding negotiation?
- Given that the authors have not built the model based upon the observed structure of negotiations and it does not seem to have any predictive power, their intentions to help negotiation organisers design a more productive process is sheer hubris. For example the authors say,"...these experiments make clear that it is relatively easy for the decision-maker in charge of organising the negotiations to manipulate the negotiation process in order to pick up the outcomes he prefers". However this is only true if the model faithfully reflects real negotiations - a fact which is, at best, completely unproved. If the authors present themselves as "experts", able to provide helpful advice, I just hope they do this from their general knowledge and not base it upon any indications from this model. History abounds with examples of "experts" persuading others that they understand phenomena that they don't.
- The conclusion has to be that the paper fails in its goals. This is because instead of choosing a model that was appropriate to the modelling target the authors chose instead to use traditional techniques which they knew did not reflect the negotiation process. Thus the paper falls between stools: it fails to produce any explanations of the outcomes in terms of the conditions; it fails to produce any new insights about agreement processes in the abstract by introducing helpful analogies for thinking about such situations; and, lastly, there is no reason to suppose it would form the basis of a successful predictive model and several reasons to suppose it would not.
- This is a pity for there is much that is good in this paper, to take a few examples: their research into relevant facts concerning the context of the negotiation process; their development of techniques to involve the stakeholders; their ideas of compositional validation; and their checking of the mechanics of their model. It is like a beautifully built car - well designed in every aspect to suit the conditions on the road, except that instead of an engine it has lead weight - however much it may impress the neighbours sitting in one's drive, it won't get you anywhere. However, all it needs is an appropriate engine and it will go like the wind.
1 There was some adjustment of these functions by the modellers to make them more appropriate to the stakeholders' goals but most of the traditional micro-economic characteristics of such functions are retained.
2 Of course, this sometimes may occur, but is not inevitable as this model suggests. Part of the problem is the extremely simple "weight" analogy for the strength of a negotiator.
3 Some of the recent biological and vetinary advice given to the UK government comes to mind. Another more pertinent example was when the previous UK government was persuaded not to lower interest rates in a recession because their economic model (wrongly) told them that a recovery was already on its way.
4 A more detailed examination of modelling methodology can be found in Edmonds (2001).
- EDMONDS, B. The Use of Models - making MABS actually work. In. Moss, S. and Davidsson, P. (eds.), (2001), Multi Agent Based Simulation, Lecture Notes in Artificial Intelligence, 1979:15-32. (http://www.cpm.mmu.ac.uk/cpmrep74.html)
Return to Contents of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 2001