Improvements and deepening for the NAMASTE model
Climate scenarios and water management policies have been tested during my thesis. However many improvements in the computer implementation are possible and could not be done during my PhD for lack of time. First to reconsider the points made in the previous paragraph, farm structure and initialization parameters are not fully scalable. Later on, the model will be able to dynamically create one or more farms from an input file.
Second, the technical management practices (ITK) are described in csv files, read by an R script that turns them into plan files mobilized by the RECORD decision plugin. For the same crop, several ITK are possible depending on the intensification of practices in terms of fertilization and irrigation. Howeverr, given the current model architecture, each ITK is also specific to the operation through the resources (labor and equipment) to be mobilized for the execution of technical operations. Thus there are currently nearly 100 plan files (5 x 10 crop intensification practices x 2 farms) to load at the beginning of each simulation which dramatically slows down the simulation process. Subsequently, the plan file creation should be dynamic so that plan files are not specific to farms (in terms of resources) and that only those mobilized for the simulation are loaded in the simulator.
The duration of technical activities is also a point on which modeling time should be spent. So far, the simulator considers that each activity takes place in one day and systematically releases resources for the next day. However, certain activities such as sowing, weeding and harvesting can take place over several days depending on the availability of the workforce. Subsequently, the operating system should be able to set the duration of a technical activity depending on the resource needs and on the available resources. It should also maintain the mobilization of resources during the whole activity duration.
To deal with the issue on soil state and plot changes, one solution may be to sub-divide a farm into smaller-sized sub-plots. Then a plot will be a combination of sub-plots and will have several STICS models running independently to simulate a crop growth.
Moreover, the management of alternative crops at a sowing or germination failures does not allow to reuse the initial rotation as the replacement crop is added to the rotation, which implies that the rotation is modified and subsequently integrates the alternative crops. However a rotation is assumed to be fixed. Even if adjustments are possible during rotation they should not be maintained for the next rotation. The management of the rotation is done at two levels: 1) at the decision level, 2) at the biophysical level.
The problem lies in the biophysical system. The rotator is a model in charge to load the subsequent simulation unit (USM) for the STICS model. At the beginning of the simulation, the same crop list is provided to the decision model and the rotator. When the replacement crop is introduced, the rotator receives an alert from the decision model to inform him about the replacement of one culture by another. The rotator will then insert this new crop in its crop list but is not able to identify this new crop as a temporary crop.
BACKGROUND ON MODELING DECISIONS IN AGRICULTURAL ECONOMICS AND AGRONOMY
Two main fields dominate decision-making approaches in farm management: agricultural economics (with bio-economic models) and agronomy (with bio-decision models) (Pearson et al. 2011). Agricultural economists are typically interested in the analysis of year-to-year strategical (sometimes tactical) decisions originating from long-term strategies (e.g., investment and technical orientation). In contrast, agronomists focus more on day-to-day farm management described in tactical decisions. The differences in temporal scale are due to the specific objective of each approach. For economists, the objective is to efficiently use scarce resources by optimizing the configuration and allocation of farm resources given farmers’ objectives and constraints in a certain production context. For agronomists, it is to organize farm practices to ensure farm production from a biophysical context (Martin et al. 2013). Agronomists identify relevant activities for a given production objective, their interdependency, what preconditions are needed to execute them and how they should be organized in time and space. Both bio-economic and bio-decision models represent farmers’ adaptive behavior.
Bio-economic models integrate both biophysical and economic components (Knowler 2002; Flichman 2011). In this approach, equations describing a farmer’s resource-management decisions are combined with those representing inputs to and outputs from agricultural activities (Janssen and van Ittersum 2007). The main goal of farm-resource allocation in time and space is to improve economic performance of farming systems, usually along with environmental performance. Bio-economic models indicate the optimal management behavior to adopt by describing agricultural activities. Agricultural activities are characterized by an enterprise and a production technology used to manage the activity. Technical coefficients represent relations between inputs and outputs by stating the amount of inputs needed to achieve a certain amount of outputs (e.g., matrix of input-output coefficients, see Janssen and van Ittersum 2007). Many farm-management decisions can be formulated as a multistage decision-making process in which farmer decision-making is characterized by a sequence of decisions made to meet farmer objectives. The time periods that divide the decision-making process are called stages and represent the moments when decisions must be made. Decision making is thus represented as a dynamic and sustained process in time (Bellman 1954; Mjelde 1986; Osman 2010). This means that at each stage, technical coefficients are updated to proceed to the next round of optimization. Three major mathematical programming techniques are commonly used to analyze and solve models of decision under uncertainty: recursive models, dynamic stochastic programming, and dynamic programming (see Miranda and Fackler 2004). Agricultural economic approaches usually assume an idealized situation for decision, in which the farmer has clearly expressed goals from the beginning and knows all the relevant alternatives and their consequences. Since the farmer’s rationality is considered to be complete, it is feasible to use the paradigm of utility maximization (Chavas et al. 2010). Simon (1950) criticized this assumption of full rationality and claimed that decision-makers do not look for the best decision but for a satisfying one given the amount of information available. This gave rise to the concept of bounded and adaptive rationality (Simon 1950; Cyert and March 1963), in which the rationality of decision-makers is limited by the information available, cognitive limitations of their minds and the finite timing of the decision. In bounded rationality, farmers tend to look for satisfactory rather than utility maximization when making relevant decisions (Kulik and Baker, 2008). From complete or bounded rationality, all bio-economic approaches are characterized by the common feature of computing a certain utility value for available options and then selecting the one with the best or satisfactory value. In applied agricultural economics, stochastic production models are more and more commonly used to represent the sequential production decisions by farmers, by specifying the production technology through a series of operational steps involving production inputs. These inputs have often the dual purpose of controlling crop yield or cattle output level on the one hand, and controlling production risk on the other (Burt 1993; Maatman et al. 2002; Ritten et al. 2010). Furthermore, sequential production decisions with risk and uncertainty can also be specified in a dynamic framework, to account for intertemporal substitutability between inputs (Fafchamps 1993). Dynamic programming models have been used as guidance tools in policy analysis and to help farmers identify irrigation strategies (Bryant et al. 1993).
Biophysical models have been investigated since the 1970s, but the difficulty in transferring simulation results to farmers and extension agents led researchers to investigate farmers’ management practices closely and develop bio-decision models (Bergez et al. 2010). A decision model, also known as a decision-making process model or farm-management model, comes from on-farm observations and extensive studies of farmers’ management practices. These studies, which show that farmers’ technical decisions are planned, led to the “model for action” concept (Matthews et al. 2002), in which decision-making processes are represented as a sequence of technical acts. Rules that describe these technical acts are organized in a decision schedule that considers sequential, iterative and adaptive processes of decisions (Aubry et al. 1998). In the 1990s, combined approaches represented farming systems as bio-decision models that link the biophysical component to a decisional component based on a set of decision rules (Aubry et al. 1998b; Attonaty et al. 1999; Bergez et al. 2006; Bergez et al. 2010). Bio-decision models describe the appropriate farm-management practice to adopt as a set of decision rules that drives the farmer’s actions over time (e.g., a vector returning a value for each time step of the simulation). Bio-decision models are designed (proactive) adaptations to possible but anticipated changes. By reviewing the decision rules, these models also describe the farmer’s reactive behavior.
Anticipated shocks in sequential decision-making processes
When decision-making process is assumed to be a succession of decisions to make, it follows that farmers are able to integrate new information about the environment at each stage and adapt to possible changes occurring between two stages. Farmers are able to anticipate all possible states of the shock (change) to which they will have to react. In 1968, Cocks stated that discrete stochastic programming (DSP) could provide solutions to sequential decision problems (Cocks 1968). DSP processes sequential decision-making problems in discrete time within a finite time horizon in which knowledge about random events changes over time (Rae 1971; Apland and Hauer 1993). During each stage, decisions are made to address risks. One refers to “embedded risk” when decisions can be divided between those initially made and those made at a later stage, once an uncertain event has occurred (Trebeck and Hardaker 1972; Hardaker 2004). The sequential and stochastic framework of the DSP can be represented as a decision tree in which nodes describe the decision stages and branches describe anticipated shocks. Considering two stages of decision, the decision-maker makes an initial decision ( 1) with uncertain knowledge of the future. After one of the states of nature of the uncertain event occurs (k), the decision-maker will adjust by making another decision ( 2 ) in the second stage, which depends on the initial decision and the state of nature k of the event. Models can become extremely large when numerous states of nature are considered; this “curse of dimensionality” is the main limitation of these models (Trebeck and Hardaker 1972; Hardaker 2004).
Flexible plan with optional paths and interchangeable activities
In manufacturing, proactive scheduling is well-suited to build protection against uncertain events into a baseline schedule (Herroelen and Leus 2004; Darnhofer and Bellon 2008). Alternative paths are considered and choices are made at the operational level while executing the plan. This type of structure has been used in agriculture as well, with flexible plans that enable decision-makers to anticipate shocks. Considering possible shocks that may occur, substitutable components, interchangeable partial plans, and optional executions are identified and introduced into the nominal plan. Depending on the context, a decision is made to perform an optional activity or to select an alternative activity or partial plan (Clouaire and Rellier 2009). Thus, two different sequences of events would most likely lead to performing two different plans. Some activities may be cancelled in one case but not in the other depending on whether they are optional or subject to a context-dependent choice (Bralts et al. 1993; Castellazzi et al. 2008; Dury et al. 2010; Castellazzi et al. 2010).
Adaptations and strategic decisions for the entire farm
Strategic decisions aim to build a long-term plan to achieve farmer production goals depending on available resources and farm structure. For instance, this plan can be represented in a model by a cropping plan that selects the crops grown on the entire farm, their surface area and their allocation within the farmland. It also offers long-term production organization, such as considering equipment acquisition and crop rotations. In the long-term, uncertain events such market price changes, climate events and sudden resource restrictions are difficult to predict, and farmers must be reactive and adapt their strategic plans.
Barbier and Bergeron (1999) used the recursive process to address price uncertainty in crop and animal production systems; the selling strategy for the herd and cropping pattern were adapted each year to deal with price uncertainty and policy intervention over 20 years. Similarly, Heidhues (1966) used a recursive approach to study the adaptation of investment and sales decisions to changes in crop prices due to policy measures. Domptail and Nuppenau (2010) adjusted in a recursive process herd size and the purchase of supplemental fodder once a year depending on the available biomass that depended directly on rainfall. In a study of a dairy-beef-sheep farm in Northern Ireland, Wallace and Moss (2002) examined the effect of possible breakdowns due to bovine spongiform encephalopathy on animal-sale and machinery-investment decisions over a seven-year period with linear programming and a recursive process.
Thus, in the operation research literature, adaptation of a strategic decision is considered a dynamic process that should be modeled via a formalism describing a reactive adaptation processes (Table 2.1).
Decision-making processes: multiple stages and sequential decisions
In Simon (1976), the concept of the decision-making process changed, and the idea of a dynamic decision-making process sustained over time through a continuous sequence of interrelated decisions (Cerf and Sebillotte 1988; Papy et al. 1988; Osman 2010) was more widely used and recognized. However, 70% of the articles reviewed focused on only one stage of the decision: adaptation at the strategic level for the entire farm or at the tactical level for the farm or plot. Some authors used formalisms such as DP and DSP to describe sequential decision-making processes. In these cases, several stages were identified when farmers have to make a decision and adapt a previous strategy to new information. Sequential representation is particularly interesting and appropriate when the author attempts to model the entire decision-making processes from strategic to tactical and operational decisions; i.e., the complete temporal and spatial dimensions of the decision and adaptation processes (see section 2.5.3). For these authors, strategic adaptations and decisions influence tactical adaptations and decisions and vice-versa. Decisions made at one of these levels may disrupt the initial organization of resource availability and competition among activities over the short term (e.g., labor availability, machinery organization, irrigation distribution) but also lead to reconsideration of long-term decisions when the cropping system requires adaptation (e.g., change in crops within the rotation, effect of the previous crop). In the current agricultural literature, these consequences on long- and short-term organization are rarely considered, even though they appear an important driver of farmers’ decision-making (Daydé et al. 2014). Combining several formalisms within an integrated model in which strategic and tactical adaptations and decisions influence each other is a good starting point for modeling adaptive behavior within farmers’ decision-making processes.
Table of contents :
Introduction – Synthesis Chapter
1.1. General introduction
1.1.1. Today’s and tomorrow’s challenges for the agriculture
1.1.2. Designing farming systems
1.1.3. Agricultural production systems and complex systems
1.1.4. Global challenges in designing agricultural production systems
1.2. Thesis project
1.2.1. Research context
1.2.2. Thesis objectives
1.3. Thesis proceedings
1.4. Thesis results
1.4.1. Literature review on adaptation in decision models
1.4.2. The Berambadi watershed
1.4.3. A methodology to guide the design of a conceptual model of farmers’ decision-making processes
1.4.4. The conceptual model NAMASTE
1.4.5. Strategic decisions on investments and cropping systems modeled with a stochastic dynamic programming approach
1.4.6. The computer model NAMASTE
1.5. Discussions and prospects
1.5.1. Review on the thesis results
1.5.2. From the demonstration prototype to the production model
1.5.3. Verification and validation
1.5.4. A decision model for the AICHA project
Processes of adaptation in farm decision-making models- A review
2.2. Background on modeling decisions in agricultural economics and agronomy
2.4. Formalisms to manage adaptive decision-making processes
2.4.1. Formalisms in proactive adaptation processes
2.4.2. Formalisms in reactive adaptation processes
2.5. Modeling adaptive decision-making processes in farming systems
2.5.1. Adaptations and strategic decisions for the entire farm
2.5.2. Adaptation and tactic decisions
2.5.3. Sequential adaptation of strategic and tactical decisions
2.6.1. Adaptation: reactive or proactive process?
2.6.2. Decision-making processes: multiple stages and sequential decisions
2.6.3. What about social sciences?
2.6.4. Uncertainty and dynamic properties
Farm typology in the Berambadi watershed (India): farming systems are determined by farm size and access to groundwater
3.2. Materials and methods
3.2.1. Case study: Hydrological and morphological description of the watershed
3.2.2. Survey design and sampling
3.2.3. Analysis method
3.3. Variability and spatialization of farm characteristics and practices
3.3.1. Farm structure
3.3.2. Farm practices
3.3.3. Water management for irrigation
3.3.4. Economic performances of the farm
3.4. Typology of farms in the Berambadi watershed
3.4.1. Characteristics of farm typology
3.4.2. Characteristics of the farm types
CMFDM: A methodology to guide the design of a conceptual model of farmers’ decision-making processes
4.2. Founding principles of the methodology
4.3. The CMFDM methodology
4.3.1. Step 1: problem definition
4.3.2. Step 2: case study selection
4.3.3. Step 3: data collection and analysis of individual case studies
4.3.4. Step 4: the generic conceptual model
4.4. Methodology implementation in a case study
4.4.1. Step 1: problem definition
4.4.2. Step 2: case study selection
4.4.3. Step 3: data collection and analysis of individual case studies
4.4.4. Step 4: the generic conceptual model
4.5. Discussion – Conclusion
Adaptive and dynamic decision-making processes: A conceptual model of production systems on Indian farms
5.2. Modeling processes
5.2.1. Indian case study
5.2.2. Modeling steps
5.2.3. Conceptual validation
5.3. A conceptual model of production systems on Indian farms
5.3.1. What should be modeled
5.3.2. Modeling in NAMASTE
5.3.3. Description of the other systems in the model
A stochastic dynamic programming approach to analyze adaptation to climate change – application to groundwater irrigation in India
6.2. Literature review on long-term farmer decisions under uncertainty
6.3.1. The farmer’s production problem
6.4. Simulations and results
NAMASTE: a dynamic model for water management at the farm level integrating strategic, tactical and operational decisions
7.2. Materials and methods
7.2.1. Conceptual modelling
7.2.2. RECORD: a modeling and simulation computer platform
7.3. Description of the farming system model
7.3.1. Models used to build the farming system model
7.3.2. Model structure
7.3.3. Dynamic functioning
7.4. Application case: the NAMASTE simulation model
7.4.1. Coupling the farming system model to the hydrological model
7.4.2. NAMASTE simulaton
7.4.3. Calibration and validation
7.4.4. Simulation results