Preamble
Applications of theories from design science to farming system design have never been explicitly reported. Still, we separate the various farming system design approaches making use of computer models into two main categories further described in the next sections (Table 1). These two categories are connected to two of the main design science theories. The first category contains the optimisation-oriented approaches. As with Simon’s theory (1969), design is mainly seen as a problem-solving process (Weersink et al. 2002). Emphasis is placed on the computational exploration of the solution space by a problem-solving algorithm. The second category contains the participatory and simulation-based approaches to farming system design. Following Schön’s theory (1983), framing the design problem is central to the design process. The subsequent exploration of the solution space to seek for alternative farming systems relies on the creativity of humans.
Table 1 Main features of the reviewed farming system design approaches making use of computer models
Optimisation approaches
Design context
The reviewed optimisation approaches were applied over a wide range of operands. Mixed farming systems, i.e. mostly dairy cattle systems relying on grasslands and crops, were most represented (Salinas et al. 1999; van Calker et al. 2004). The plant farming systems considered were as diverse as cherries (Cittadini et al. 2008), bulbs (Rossing et al. 1997) and vegetables (Dogliotti et al. 2005). Whereas some of the approaches were confined to a particular type of farming system (Cittadini et al. 2008; Rossing et al. 1997), others were applicable to several types of system, e.g. vegetables, beef cattle and mixed production systems (Dogliotti et al. 2005) or arable, beef and dairy cattle, sheep and goats, and mixed production systems (Louhichi et al. 2010).
In most reviewed approaches, the problem raised by the farming system was very specific and well known, e.g. modification of the common agricultural policy (Veysset et al. 2005), environmental side effects related to high consumption of pesticides and fertilizers (Rossing et al. 1997) or inefficient use of available farm resources (Salinas et al. 1999). In the remaining cases, the nature of the problem remained open and had to be clarified for each application (Bernet et al. 2001; Castelan-Ortega et al. 2003; Groot et al. 2007; Louhichi et al. 2010). Following Simon’s view, the problem was assumed to be stable in all cases. The single source of change found did not concern the nature of external drivers but their variability over time, e.g. variability of weather (Cabrera et al. 2006; Flaten and Lien 2007) and of output prices (Mosnier et al. 2009).
Except for one approach (Rossing et al. 2009b), the operators addressed the design problem on the sole basis of their background knowledge, that is, without addition of new knowledge either imported or learned during the design process. This problem was in all cases considered to be an optimisation of the configuration of farm resources and their allocation over space and time given farmers’ production objectives, constraints, production context (current or modified) and farm resources (current or modified). As a consequence, the definition of the solution space to be explored was determined by the variety of strategic, and in a few cases (Flaten and Lien 2007; Woodward et al. 1995), the tactical decisions considered. For instance, the number of possible crop rotations to be selected for allocation to land units (Dogliotti et al. 2005) or the number of sets of management practices (i.e. a diet and a confinement time for cows and a crop rotation; Cabrera et al. 2006) conditioned the size of the solution space.
The primary concern behind a change in the configuration of farm resources and their allocation over space and time was the improvement of economic performance of existing farming systems. This goal was represented explicitly as a gross margin at the farm scale (Veysset et al. 2005) or per hectare, or implicitly through agronomic factors affecting economic performance, e.g. an improvement of herbage use efficiency at grazing (Woodward et al. 1995). Improvement of economic performance was generally associated with environmental and/or social goals. Environmental goals were clearly formulated regarding, for instance, soil erosion rate (Dogliotti et al. 2005), nitrogen loss, plant species number (Groot et al. 2007) and CO2 emissions (Ramsden and Gibbons 2009). Social goals were not as straightforward but the workload of the farming system designed was often taken into account in the analysis (Cittadini et al. 2008).
Intended users of the representations of farming systems resulting from optimisation varied greatly among the reviewed approaches. In all cases, researchers were primary users of the optimisation results that appeared in scientific publications. In some cases, farmers (Rossing et al. 2009b; Veysset et al. 2005), farm advisors (Bernet et al. 2001), policy-makers (Louhichi et al. 2010; van de Ven and van Keulen 2007) or a range of them (farmers, farm advisors, experts in Cabrera et al. 2008) were other intended users.
Design process
There was not much participation of intended users or other stakeholders in the design process. In about half of the reviewed optimisation approaches, there was none (Flaten and Lien 2007; Louhichi et al. 2010; van de Ven and van Keulen 2007). Consequently in most such approaches, and as already observed with Simon’s theory, the problem raised by the farming system was assumed to be clear and almost taken for granted (Fig. 4). The decisions to be made, the ends to be achieved and the means chosen were extracted by research on practical situations. The translation of goals, criteria, constraints and alternatives into a solution space was thus carried out by researchers. The three cases of participation during analysis and conceptualization of the problem situation were collaborative, with the involvement of stakeholders during definition of design goals and identification of management constraints (Rossing et al. 1997, 2009b; van Calker et al. 2006).
In line with Simon’s theory, most effort was then invested in the problem-solving step, i.e. the generation of a solution by computational exploration of the solution space with problem-solving algorithms (Fig. 4). Such optimisation procedures consisted of the selection of a number of activities (e.g. cropping, grass production for grazing, herd management; Salinas et al. 1999) out of a set defined in an input–output matrix. The development of the set of activities was either performed “by hand” (Berentsen and Giesen 1995) possibly based on consultative participation of stakeholders through farm surveys (Castelan-Ortega et al. 2003) or with computer models (Cittadini et al. 2006). For instance, Dogliotti et al. (2003) developed a crop rotation generator that combines crops from a predefined list to produce all possible rotations, given a number of agronomic filters related, for example, to undesirable crop successions. Quantification of inputs and outputs for each activity used consultative participation of experts (Cittadini et al. 2006; Dogliotti et al. 2004) and computer models. In one case (Cabrera et al. 2008), a collegiate participation of stakeholders was required to adapt these computer models to the design purpose.
Once the input–output matrix of activities was available, the optimisation procedure was run (Fig. 4). While maximising agronomic, economic, environmental and/or social goals once (Cittadini et al. 2008; Louhichi et al. 2010) or several times (Flaten and Lien 2007; Mosnier et al. 2009), problem-solving algorithms selected and allocated activities to a farming system possibly divided into subunits differing, for example in terms of soil type. This optimisation was subject to constraints at the farm level to preserve the consistency of the designed system. For instance, in livestock farming systems, selection and allocation of activities had to consider that feeding requirements had to match on-farm feed production and purchased feed (Berentsen and Giesen 1995; Veysset et al. 2005).
Practical evaluation of the solution identified was rare, as was the possible feedback between problem definition, generation and evaluation of this solution. In two cases only (Rossing et al. 1997, 2009a), the authors reported collaborative participation of stakeholders to discuss the extent to which the solution identified matched the design goals.
Computer model structure
Except for Woodward et al. (1995), whose model operated at the grazing system scale, that is the set of grassland fields grazed and the grazing animals, optimisation models used for generating alternative farming systems have always operated at the farm scale (Cabrera et al. 2006). The smallest spatial modelling units were so-called land units, i.e. areas of land uniform as regards soil and climatic conditions and the farmer’s management practices (Cittadini et al. 2008; Dogliotti et al. 2005; Rossing et al. 1997). In most cases, such land units aggregated several fields. The models of Groot et al. (2007) and Woodward et al. (1995) were the sole optimisation models representing fields individually and even paddocks for the latter. For animals, the smallest modelling unit was an average animal, representative of the herd or a part of it (Bernet et al. 2001; Herrero et al. 1999). At these smallest scales, agricultural activities (crop rotations, temporary or permanent grasslands, dairy cows, etc.; Janssen and van Ittersum 2007) were described by technical coefficients defining the amount of inputs required to achieve a certain level of outputs (agronomic, environmental, economic and/or social; van Ittersum and Rabbinge 1997).
Input–output combinations were calculated from mechanistic and empirical crop, grassland and animal models operating on a daily time step (Castelan-Ortega et al. 2003; Louhichi et al. 2010; Mosnier et al. 2009), or alternatively deduced from experiments (Nielsen et al. 2004; van de Ven and van Keulen 2007), farm data (Veysset et al. 2005) or expert knowledge (Cittadini et al. 2008; Dogliotti et al. 2005), especially when innovative or poorly informed alternatives to the current agricultural activities were considered. At the farm scale, most reviewed optimisation models were static. In such cases, for instance in the case of crop rotations, technical coefficients used as inputs to the optimisation models were averages of input–output combinations over several years. The remaining models were dynamic optimisation models (Cabrera et al. 2006; Mosnier et al. 2009) proceeding in a succession of optimisation rounds, often with time steps of 1 month. This meant that every month, technical coefficients were updated to proceed to the next round of optimisation. In such cases, the time horizon of the models ranged from one (Flaten and Lien 2007) to several (Mosnier et al. 2009) years. Consequently, while models operating over 1 year considered weather as the single variable environmental factor affecting the behaviour of the modelled system (Woodward et al. 1995), models operating over several years also included market conditions (Mosnier et al. 2009).
With the exception of the model of Woodward et al. (1995) which seeks to maximise animal intake at grazing, all the reviewed models allocated agricultural activities to maximize economic return under constraints related, among other things, to farm resources, environmental performance (soil erosion, soil organic matter depletion, nitrogen leaching, etc.), and overall consistency of the system. Farm management was always treated as an optimisation summarized in an objective function. The optimisation procedure was performed using various mathematical techniques, i.e. linear programming (Dogliotti et al. 2005), stochastic programming (Flaten and Lien 2007), dynamic recursive optimisation (Mosnier et al. 2009), branch and bound optimisation (Cabrera et al. 2006) and multi-objective genetic algorithms (Matthews et al. 2006a).
Participatory and simulation-based approaches
Design context
As with optimisation approaches, operands of the reviewed participatory and simulation-based approaches were mostly livestock farming systems. These were either grassland-based beef or dairy cattle farming systems (Andrieu et al. 2007; Cacho et al. 1995; Romera et al. 2004) or mixed beef and dairy cattle farming systems relying on grasslands and crops (Rotz et al. 2009; Vayssières et al. 2009). Smallholder mixed-farming systems included more marginal crops and livestock species such as quinoa, beans and donkeys (Herve et al. 2002; Pfister et al. 2005). Applications to plant farming systems considered banana (Blazy et al. 2010) and arable crops (Keating et al. 2003).
With participatory and simulation-based approaches the problem raised by the farming system was not always clearly formulated in the scientific articles. Exceptions concerned for instance adaptation to the CAP reform (Matthews et al. 2006b), biodiversity conservation (Jouven and Baumont 2008) and instability of production due to variability of environmental factors such as weather (Martin et al. 2011b; Romera et al. 2004). Two reasons explain this loose formulation of the design problem. First, few approaches (Rivington et al. 2007) addressed design problems considering long term horizon. In such cases, in line with Schön’s theory, uncertainty about the problem situation is high. Second, most reviewed articles focused on the computer models used in the course of the design process. The function, behaviour and structure of these models were then presented in details at the expense of the wider design context and problem situation specific to a given application. A problem pointed out by most participatory and simulation-based approaches was the need to improve the understanding of farming systems, their dynamics and internal interactions in particular (Cacho et al. 1995; Keating et al. 2003).
In most cases, the background knowledge of the operators did not change much during the design process. However, in contrast with the optimisation approach, a significant part of the knowledge remained in the mind of the operators and was applied only during the design process instead of being incorporated in the computer model. Still, approaches focused on long-term problems had no preconceived idea about solutions. Changes in the configuration of farm resources and their allocation over space and time by farmers at the strategic, tactical and/or operational levels were then explored. For instance, Kerr et al. (1999a, b) examined changes in land use affecting the whole farming system while Cros et al. (2004) mainly dealt with tactical and operational changes in the feeding and grazing management of dairy cows. One noticeable difference with optimisation approaches is that the reviewed articles (Pfister et al. 2005; van Wijk et al. 2009) pointed to the need to improve our understanding of current farming systems to seek for relevant solutions to their problems. For instance, it was felt necessary to characterize the impact of climate change on farming systems before tackling the identification of possible adaptation strategies (Rivington et al. 2007; Martin et al. 2011a).
The reviewed approaches mostly aimed at improving the agronomic performance of the farming systems. The corresponding goals were yields (Keating et al. 2003; Tittonell et al. 2010) or herbage use efficiency at grazing (Andrieu et al. 2007; Romera et al. 2004). In some cases, multiple goals were examined. Economic performance was assessed for instance with a net margin per hectare (Blazy et al. 2010). Environmental goals for instance included grassland biodiversity (Jouven and Baumont 2008), pesticide use (Blazy et al. 2010) and greenhouse gas emissions (White et al. 2010). Social goals focused on the workload to ensure the feasibility of the farming systems designed (Herve et al. 2002; Martin et al. 2011b).
Again, intended users of the representations of alternative farming systems produced were in all cases researchers as users of such representations for publishing their findings. Farmers (Vayssières et al. 2009), farm advisors (Martin et al. 2011b), teachers (Machado et al. 2010) or a range of them (farmers, bank managers, loan officers and farm advisors in Kerr et al. 1999a, b) were also involved in the design process.
Design process
Participation of stakeholders in the design process was much greater with participatory and simulation-based approaches. It was nil for only 6 over the 20 approaches considered. Thus from the very early stages of the design process, stakeholders were involved in analysing the problem situation (Fig. 5). The most extreme case was Vayssières et al. (2009) where the main researcher was immersed in the daily life of farmers and participated in the farming activities in order to gain credibility in the eyes of farmers and to assess with them what were their most critical problems. Carberry et al. (2002) pursued the same objective with a collegiate participation as well. This position follows Schön’s point of view, as the invitation for stakeholders’ participation indicates that researchers consider they do not have adequate descriptions of the problem, of perceived solutions and goals and of the promising farming systems reachable from the current farming systems. In addition, computer models were used at this stage in a few of the reviewed approaches (Martin et al. 2011a; Rivington et al. 2007). They enabled problem situation analysis to characterize the exposure and vulnerability of farming systems to climate change. Climate models and crop and grassland models were used to characterize the local impact of climate change on crop and grassland production.
In contrast to optimisation approaches, exploration of the solution space to generate a solution relied on the creativity of researchers and involved stakeholders (Fig. 5). As with Schön’s Reflection-in-Action paradigm, this exploration consisted of identifying “by hand” an initial possible solution and in defining contextualised simulation experiments, possibly through collegiate participation (Carberry et al. 2002; Martin et al. 2011a; Matthews et al. 2006a) to evaluate the relevance of this solution in order to return to the problem definition and to correct the initial solution. At this stage, computer models could be used to develop specific artefacts supporting the design process (Fig. 5). For instance, Martin et al. (2011a) have developed flattened wooden sticks (that they call forage sticks) marked with the forage yield in kilograms per hectare and per 4-week period across the calendar year for a number of forage crop including new crops in the area, and its year-round management (e.g. early and productive permanent grassland grazed six times a year) including novel practices. During participatory workshops, players had to select sticks and to assign an area to each selected stick as a first step towards the design of a whole livestock production system. Such forage sticks have been built using crop and grassland models.
Using a computer model (Fig. 5), evaluation of alternative farming systems was conducted through virtual experimentation, reproducing their behaviour over a given period and under different production contexts (Romera et al. 2004; Vayssières et al. 2009). For instance, in Martin et al. (2011a), after players had selected forage sticks and assigned an area to each selected stick, they used a balance model to evaluate whether forage production in the assemblage created by the players would match animal feeding requirements. On the basis of the output, players were invited to reconsider their tentative solution (Fig. 5). The evaluation stage also enabled consultative participation of farmers through discussions about the relevance and feasibility of alternative farming systems (Martin et al. 2011b).
Throughout the design and evaluation loops, the model user(s) and involved stakeholders proceeded empirically by trial and error to explore the solution space and progressively elaborate a satisfactory solution that achieved the desired goals whilst satisfying constraints (Fig. 5). For instance, Carberry et al. (2002) described the behaviour of farmers exploring their own system and learning through the use of a computer model, rather like “learning by doing”. Similar observations are reported in Duru and Martin-Clouaire (2011).
Computer model structure
The mechanistic computer models used for problem situation analysis simulated dynamically crop and grassland production at the field scale on a daily time step and over a single production cycle (Martin et al. 2011a; Rivington et al. 2007). The single environmental factor considered was weather, to characterize the local impact of climate change on crops and grasslands. Such models focused mainly on biophysical aspects, with detailed representations of the soil and crop/grassland components of the field. However, most of them paid little attention to the representation of farmers’ decision-making processes. In the simplest case, farm management was seen as a sequence of technical actions on fixed dates as simulation inputs (Stöckle et al. 2003 used in Rivington et al. 2007). A more elaborate representation, i.e. the rule-based approach, was used in Martin et al. (2011a). It relates the decisions made and related actions to conditions encountered dynamically. The computer models used for problem situation analysis thus corresponded to so-called crop and grassland models.
Simulation models enabling evaluation of alternative farming systems all operated at the farm scale (Jouven and Baumont 2008; Shaffer et al. 2000), except for a grazing system model (Cros et al. 2003). The smallest spatial modelling unit was generally a field (Romera et al. 2004; Pfister et al. 2005). In a few cases, it was the so-called land unit, possibly aggregating several fields (Blazy et al. 2010; Herve et al. 2002) and in one case, the whole farm was treated as if it was a single paddock of pasture (White et al. 2010). Animal processes were modelled at the scale of individual cows (Romera et al. 2004), of average representative animals (Andrieu et al. 2007; Martin et al. 2011b), of the whole herd (White et al. 2010).
The biophysical models integrated within the farm models were either mechanistic (Keating et al. 2003; Shaffer et al. 2000), statistical (Herve et al. 2002; Pfister et al. 2005), mechanistic and statistical (Cacho et al. 1995; Kerr et al. 1999a, b) or mechanistic and empirical (Martin et al. 2011b). The biophysical models were either static (White et al. 2010) or dynamic, operating on a time step of 1 year (Kerr et al. 1999a, b), 1 week (Blazy et al. 2010) or 1 day (Keating et al. 2003), even when the related farm model was static (Martin et al. 2011a). The remaining dynamic farm models had a time horizon ranging from one season (Cros et al. 2003) to several years (Blazy et al. 2010; Cacho et al. 1995). Weather was the main environmental factor influencing the behaviour of the modelled system (Andrieu et al. 2007), but market conditions and policy context were considered in some cases (Blazy et al. 2010; Matthews et al. 2006b).
Farm management was generally modelled with rule-based systems (Cros et al. 2003). However, this approach offers no powerful means to create links between rules so as to control the order in which they are used. Hence Martin et al. (2011b) used a recent and more sophisticated approach (Martin-Clouaire and Rellier 2009) called activity-based models, in which management is seen as the problem of coordinating activities because these require resources which are either limited or constrained by availability over time and also because future activities need to be anticipated in relation to present ones. Instead of using rules, management is thus represented as a set of activities organised in plans that are flexible and adaptable to changing conditions. In contrast, a few models neglect the dynamics of farm management by assuming fixed strategic and tactical decisions defined as inputs (Kerr et al. 1999a, b; Pfister et al. 2005). Mathematical techniques used to integrate the behaviour of the simulated system ranged from continuous time simulation (Shaffer et al. 2000), to discrete event simulation (Martin et al. 2011b), stock-flow simulation (Pfister et al. 2005; Vayssières et al. 2009), spreadsheet models consisting of static balances (Martin et al. 2011a) and a combination of static models (White et al. 2010).
Connection and differentiation
Clearly the optimisation approach corresponds to Simon’s problem-solving theory. The participatory and simulation-based approach shares with Schön’s theory the emphasis on the key roles played by stakeholders’ participation and tacit knowledge. Because Schön’s theory provides principles rather than practical methods, it does not give a specific status to simulation. The fundamental differences between Simon’s and Schön’s theories remained when comparing the two approaches used in farming system design. These differences were concerned with several aspects including the importance given to each step of the design process, the participation of stakeholders to this process, the function, behaviour and structure of the computer models used, and the type of interactions between the model and the stakeholders involved (Figs. 4 and 5).
With the optimisation approach, the computer model includes (a) a representation of the relations (constraints) that the system to be designed (e.g. a farm) should satisfy and (b) the aggregated criteria that enable one to discriminate between several potential design solutions. In the participatory and simulation-based approach, the computer model represents only what the system to be designed is composed of and how these components interact dynamically with the external environments of the system and internally between them. In the participatory and simulation-based approach, the criteria are not formalized; they remain in the heads of the operators, who may have different values and preferences.
In the optimisation approach, the computer model of the system is typically static (a set of algebraic equations in which time is not explicitly represented) whereas in the participatory and simulation-based approach, the computer model of the system is dynamic and responsive to external factors, e.g. weather. Dynamic models keep changing with reference to time whereas static models are at equilibrium in a steady state; the equilibrium might change but not necessarily in relation to a notion of time advancing.
In the optimisation approach, the computer model of the system is essentially a constraint enforcement framework that fully delimits the solution space. In the participatory and simulation-based approach, the computer model is not imposing such a strong restriction on the solution space. In addition, the operators have much more freedom to change the computer model than with the optimisation approach because the latter requires computer models that should be simple enough in order to remain computationally tractable by the optimisation algorithm.
The modelling effort in the optimisation approach is different than in the participatory and simulation-based approach because the representation is done at a very abstract level (e.g. numerical variables linked by linear equations) required by the optimisation algorithm. Consequently modifying the computer model in the design process requires a highly technical competence and operators that do not possess it cannot contribute. In the participatory and simulation-based approach, the mapping between reality and representation is more direct (the concepts used to characterise the reality have their direct counterpart in the simulation model). For this reason, it is much easier for the operators to pinpoint the reasons for particular behaviour of the computer model (causality can be traced) and to suggest changes to the computer model (i.e. changes to the model of the system to be designed).
Finally, the optimisation approach and participatory and simulation-based approach differ in their merits and limitations: the optimisation approach takes advantage of the computational power of machines to explore efficiently a large solution space, the main limitation coming from representation restriction imposed by the optimisation machinery. The participatory and simulation-based approach emphasises temporal-based thinking, that is, thinking about how the system and its environment may evolve over time, which makes it possible to evaluate situational decision-making options. In the participatory and simulation-based approach, the main limitation relates to the absence of optimisation: a much better solution might lie not very far from the one produced by the participatory and simulation-based approach.
In spite of the above differences, the reasons for choosing one particular approach over another are seldom addressed in the scientific literature. Apart from the author(s)’ allegiance to a particular scientific community and its idiosyncrasies, one could relate this choice to the applications considered, which can best be described by the design context and by the nature of the innovations developed for farming systems to cope with the changing world. If differences in problem formulation were evident, the nature of the problems tackled, the kind of decisions taken by intended users (e.g. national vs. on-farm policy) and the wider design context were in the end not fundamentally different. The nature of the innovations generated by the two categories of farming system design approaches making use of computer models is further analysed in the next subsection.