Sie sind auf Seite 1von 16

Mathematics and Computers in Simulation 68 (2005) 339354

Optimal harvesting strategies for a multi-cycle and multi-pond shrimp operation: A practical network model
Run Yu a,b , PingSun Leung b,
b

Department of Economics, University of Hawaii at Manoa, Honolulu, HI 96822, USA Department of Molecular Biosciences and Bioengineering, University of Hawaii at Manoa, 3050 Maile Way, Gilmore 111, Honolulu, HI 96822, USA

Received 18 June 2004; received in revised form 20 January 2005; accepted 31 January 2005 Available online 7 April 2005

Abstract In this paper, we introduce a network formulation of the optimal scheduling model for a multi-cycle and multipond shrimp operation grounded on the original optimal harvesting theory for a single production unit. The optimal schedule comprises the harvesting and restocking time that maximizes total prot throughout the planning horizon, bounded by the underlying biological and economic conditions. The model takes into account the information such as harvest size distribution, seasonality of price, temperature and weight-dependent growth, and labor force and market demand constraints. We applied the model to an existing shrimp operation in Hawaii with 40 one-acre ponds and generated the optimal schedule for a year that maximizes overall production. The model schedule is found to be able to increase total production by 5% when compared to the schedule generated using an educated trial-and-error procedure currently practiced by this operation. Further insights for this multi-cycle and multi-pond scheduling problem were also generated through several alternate simulations. It is found that labor force and market demand constraints are the key factors in scheduling multi-cycle and multi-pond shrimp operations. 2005 IMACS. Published by Elsevier B.V. All rights reserved.
Keywords: Optimal production schedule; Shrimp farming; Aquaculture; Network model

Corresponding author. Tel.: +1 808 9568562; fax: +1 808 9569269. E-mail addresses: run@hawaii.edu (R. Yu), psleung@hawaii.edu (P. Leung).

0378-4754/$30.00 2005 IMACS. Published by Elsevier B.V. All rights reserved. doi:10.1016/j.matcom.2005.01.018

340

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

1. Introduction The problem of determining the optimal harvest size of farmed sh has rst been analyzed and modeled by Bjorndal [2] using an optimal control framework for salmon culture. Arnason [1], Heaps [8,9], and Hean [7] have extended this model to include optimal feeding schedule, density-dependent growth and mortality and potential culling, and release cost, respectively. Cacho et al. [3] have applied a similar analytical framework to identify the feeding strategies of catsh culture. Springborn et al. [18] have analyzed the optimum harvest time for cultured Nile tilapia using a similar framework. Leung et al. [14] have also employed a similar framework to identify the optimal harvest age for giant clam culture. Recently, Pascoe et al. [15] have applied a similar framework to identify the optimal harvest time for sea bream and tiger prawn. Forsberg [4,5] has proposed an alternative multi-period linear programming approach to model explicitly the size-structure of farmed sh in determining the optimal stocking and harvesting schedules. In terms of applications to crustaceans, Leung and Shang [12] have applied Markov process and dynamic programming to model the stocking and harvesting decisions for freshwater prawns. Karp et al. [11] have also applied dynamic programming to determine the optimal stocking and harvesting rates of shrimp in the Southwest U.S. Hochman et al. [10] and Leung et al. [13] have extended the stochastic growing inventory framework to marine shrimp culture for the purpose of identifying the optimal stocking and harvesting schedules. Tian et al. [20] have evaluated the implications of using different shrimp growth functions on optimal harvest size. Purwanto [16] has developed a set of linear and non-linear programming models for optimizing the economic benets from shrimp farming in Australia and West Java, Indonesia. Tian et al. [19] have also developed a computer simulation model to analyze the economics of shrimp production under different stocking regimes, harvesting schedules and farm sizes. While the previous research mentioned above has provided the theoretical foundation to model optimal harvesting decisions for nsh and shrimp culture, they are generally not readily applicable to commercial operations for the following reasons. First, most of the models described above used experimental data to model the growth process and they have been shown to differ widely from actual commercial operations. For example, the model by Leung et al. [13] was based on limited data from one growout cycle of marine shrimp in an experimental pond. Second, all of the models are designed primarily as a research inquiry tool rather than an operations management tool for on-farm applications. This is particularly true for the model by Leung et al. [13] where it has been used primarily to investigate the economics of controlled environment. Third, almost all of the models were developed for optimizing the harvest schedule of a single pond or production unit. While it is certainly important to derive and understand the essence of the optimal harvest schedule of a single pond, it is essential to extend this framework into a whole farm setting whereby all the ponds or production units are scheduled in an optimal manner to smooth out production so as to satisfy the labor force and market demand constraints. Against this background, the purpose of this research is to develop, test and operationalize a management model that takes into account the multi-cycle and multi-pond nature of most commercial shrimp farms in order to solve for the optimal growout production schedule of the entire farm. The paper is organized as follows: Section 2 presents the mathematical representation of the optimal scheduling problem and introduces the underlying biological and economic functions. Section 3 provides the network representation of the optimal scheduling model that is equivalent to the otherwise unsolvable mathematical formulation. Section 4 applies the network model to an existing shrimp farming entity with 40 growout ponds to generate the optimal schedule for a year. Section 5 discusses the results of

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

341

the optimal scheduling and implications from several alternate simulations. Section 6 concludes the paper.

2. The mathematical formulation In this section, we present the mathematical formulation of the optimal scheduling problem and explain the underlying biological and economic functions involved. Our whole-farm optimal scheduling model is grounded on the optimal harvesting theory pioneered by Bjorndal [2]. The problem for the planner at hand is to determine the schedule of harvesting and restocking time that maximizes total prot for a multi-pond shrimp operation for a nite planning horizon that can accommodate several production cycles. In order to provide insight into this problem, we start analyzing our whole-farm optimal scheduling problem by considering a single representative pond. The nite planning period is continuously numbered in the time decision unit, e.g. a day or a week, or t = 1, 2, 3, . . ., T. T is the ending time when all shrimps are assumed to be harvested. Dene T = {t|t = 1, 2, 3, . . . , T }. The schedule for a single growout production cycle can be represented by a vector h(i, j), i, j T. h(i, j) denes a single growout production cycle that begins at time i and ends at time j. Dene the set h = {h(i, j)|i, j T; i < j}. h is composed of all feasible schedule for a single production cycle. As a result, the multi-cycle production schedule for the whole planning period can be constructed as the combination of elements in h that does not conict with the time path. Let Hl be such a combination that consists of m elements (h1 (i1 , j1 ), h2 (i2 , j2 ), . . ., hm (im , jm )) from the set h. Rank these vectors according to their rst element i in the bracket from the smallest to the largest, the non-conict time path condition requires that 1 i1 < j1 < i2 < j2 < < im < jm T . Following the denition of h(i, j), is and js represent the planting and harvesting time, respectively. The non-conict time path condition simply ensures that production cycle cannot overlap each other. Therefore, the feasible multi-cycle schedule set can be dened as follows: H = {H(h1 (i1 , j1 ), h2 (i2 , j2 ), . . . , hm (im , jm ))|h(i, j) h, 1 i1 < j1 < i2 < j2 < < im < jm T, m N+ } Each element in H represents a feasible multi-cycle schedule that covers the whole planning period. For example, H(h(1, 100), h(105, 230), h(T 100, T )) denotes a management schedule that consists of three production cycles that begin at time 1, 105, and T 100 and end at time 100, 230, and T, respectively. Now we proceed to construct the planners objective function through introducing the underlying specic biological and economic functions. 2.1. The growth function The fundamental factor in production scheduling is the growth function. As many studies suggested, the incremental shrimp weight can be modeled as a function of various variables such as feeding, stocking size, biomass, density and temperature [4,15,17,21]. In practice, if we assume that controllable variables such as feeding, biomass, and stocking density are already at the optimal levels, the incremental shrimp

342

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

weight can be treated as if it only depends on temperature and the previous periods weight that can be expressed as follows: Wi,t+1 = Wi,t + G(Tt , Wi,t ) (1)

where W is the average weight of farmed shrimp measured at the beginning of each discrete time; its rst subscript i denotes the planting time and the second subscript t denotes current time; G(Tt , Wi,t ) is the growth increment during time t to t + 1. Therefore, expression (1) states that the current average shrimp weight equals the average shrimp weight at the preceding time plus the growth increment. As a result, for a feasible production cycle schedule h(i, j), the average weight of the harvested shrimp is equal to
j1

Wi,j = Si +
t=i

G(Tt , Wi,t )

(2)

where Si is the average shrimp weight when they are planted at time i; Si S, or S = {Si | i T}. The elements in S are the average weight of shrimp available for planting at each discrete time. 2.2. The survival function Similar to the growth function, the number of surviving shrimp can be expressed as Ni,j = Ni,i Mi,j (Bi,j ) (3)

where Ni,j is the number of surviving shrimp at harvesting time j; Ni,i is the number of shrimp at planting time i. The survival rate Mi,j is a function of biomass, Bi,j and is assumed to decrease with an increase in biomass. 2.3. Size distribution function and price function The above growth function actually focuses on the average weight of a shrimp cohort. However, the actual harvested shrimps vary in size. Moreover, larger size of shrimp commands a higher price. As a result, harvest size distribution has an important impact on optimal scheduling [19]. Furthermore, in many situations, price is characterized by seasonality [14]. In order to capture these effects, size distribution function and price function are introduced, respectively. Distribution function k (Wi,t ) is the percent of t total harvested shrimp number that falls into the kth size category. It is a function of the average shrimp weight. The parameter k represents the category of shrimp size, e.g. size counts of 21/25, 26/30, 31/35 and 36/40. The corresponding price function for the kth size category is Ptk (k, t). Basically, it is a timedependent function and thus captures the seasonality effect. The price for larger size shrimp is assumed to be always higher during a particular season. 2.4. Cost function The total production cost is composed of variable costs and xed costs and is expressed as follows:
j1

Ci,j =
t=i

Ni,t Pf,t Et G(Tt , Wi,t ) + OVC(i, j) + FC(i, j)

(4)

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

343

where the rst component on the right hand side is the feed cost function; Pf,t is the unit price for feeds at time t; Et is the food conversion efciency at time t, which denotes the units of feeds required for one unit of growth. OVC(i, j) represents variable costs other than feed costs and FC(i, j) is the xed cost for one production cycle. The feed price Pf can be modeled as a time-dependent function to take into account seasonality of price. Also, the food conversion efciency Et can be modeled as a function of weight and temperature. 2.5. Market value function and prot function We now turn to the derivations of the market value for every shrimp size category and the prot for a single harvest, which can respectively be expressed as Eqs. (5) and (6) as follows:
k k Vi,j = Pj Ni,j k (Wi,j ), j K

k = 1, 2, . . . , K

(5) (6)

Vi,j =
k=1

k Vi,j Ci,j

Finally, total prot for a feasible multi-cycle schedule Hm can be expressed as NVHm =
h(i,j) Hm

(Vi,j Ci,j ),

Hm H

(7)

We now proceed to the scheduling of several production ponds simultaneously. The multi-pond scheduling will not differ from single pond scheduling in nature, if there is no constraint on the frequency of harvesting and restocking. In that situation, the optimal scheduling can be obtained by applying the single pond scheduling model to each pond respectively. However, the multi-pond scheduling problem becomes complicated and different from sequentially solving the single pond scheduling problem when the frequency of harvesting and restocking are restricted due to the limitation of labor force and market demand. 2.6. The labor force and market demand constraints In practice, due to the limitation of labor force, facility capacity and market demand, shrimp farmers cannot harvest shrimp unlimitedly at any one time. Furthermore, shrimp farmers would also tend to spread their production throughout the year in order to better utilize the available labor force as well as to smooth out their supply to meet market demand and cash ows. We introduce these constraints into the model simply in terms of minimum and maximum frequency of harvesting and restocking. The constraints can be expressed as follows:
N

for shrimp planting at time i

CS min
n=1

hn (i, j) CS max
N

(8)

for shrimp harvesting at time j

CH min
n=1

hn (i, j) CH max

(9)

344

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

where hn (i, j) denotes a production cycle vector for the nth pond which begins at time i and ends at time j. CS min , CS max , CH min and CH max are constants depicting the maximum and minimum number of plantings and harvestings. Expression (8) shows that at any one time, there must have at least CS min plantings but not to exceed CS max plantings. Similarly, expression (9) ensures that at least CH min harvestings to be performed at any one time but there can be no more than CH max harvestings. 2.7. The optimal scheduling problem As stated before, the objective of the planner is to maximize prot from production of all ponds throughout the entire planning period. Therefore, the multi-cycle and multi-pond optimal scheduling problem can simply be viewed as choosing one element from each ponds feasible multi-cycle scheduling set Hn such that their combinations can result in maximum prot under conditions restricted by expressions (8) and (9). The overall maximization problem can now be expressed as
N K k=1

Max
n=1 H H n N

k Vn,i,j Cn,i,j

CS min s.t. CH min


n=1 n=1 N

hn (i, j) CS max for i T

(10)

hn (i, j) CH max for j T

In order for the present model to be of practical use, we will have to transform the mathematical model into an available commercial solver in order to solve the optimal scheduling problem. However, we will encounter two major difculties inherent in this mathematical formulation. One difculty is the intrinsic nonlinearity of the various underlying functions such as the growth function and the harvest size distribution function. Even if we can overcome this nonlinearity issue by either linearizing the nonlinear functions or simply adopting a nonlinear programming framework, we will still be unable to solve this inherently difcult 01 integer programming formulation. Suppose we let a 01 binary variable to represent the management decision where 1 denotes harvesting/restocking to occur and 0 denotes otherwise. Using that representation, the feasible scheduling set will consist of 2NT elements for a scheduling problem for N ponds and T time units. For example, even solving a 1-year weekly planning of 10 ponds will be prohibitive in terms of computing time even using the most efcient branch and bound technique and on the fastest computer. This well-known NP hard problem of integer programming has rendered such formulation to be impractical. On the other hand, the network-related construction has been well documented to be an efcient approach to handle this kind of integer discrete problem [6]. Therefore, we will turn to the network representation of our optimal scheduling model for a practical solution strategy. 3. The network representation As we have done previously, we will introduce how to convert our mathematical formulation into a network representation by rst considering a representative pond.

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

345

Fig. 1. Network structure of a representative pond.

Fig. 1 illustrates the network structure for a representative pond. The node (circle) represents one discrete time in the planning horizon when a managerial decision can take place. Node 0 denotes the time when the planner prepares the schedule. The series of nodes from 1 to T represents the set T. The arc from each node to each later node represents a potential single production cycle schedule where the starting node denotes the planting time and the ending node denotes the harvesting time. As a result, each arc corresponds to an element in the set h. Furthermore, a feasible multi-cycle management schedule is the connection of arcs that does not overlap each other. For example, in Fig. 1, the connection of arcs a, d, g and j and the connection of arcs a, e and i are feasible multi-cycle management schedules, while the connection of arcs b, d, g and j and the connection of arcs a, e, g and j are not feasible, because arc b overlaps arc d and arc e overlaps arc g.

Fig. 2. Representation of network structure in a decision tree form.

346

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

Fig. 3. An example of a feasible multi-cycle schedule. Node conservation requirement: the ow on every arc must be either 1 or 0. Flow conservation requirement: sum of inows equals sum of outows for every node. Note: the ow is counted as inow for ending node, while counted as outow for beginning node. Hence, the status of inow represents the harvesting decision and the status of outow represents the restocking decision.

Fig. 2 illustrates the networks managerial process in the form of a decision tree where the ith node has (T i) branches. At each node, we will choose one and only one branch. NV(i, j)s are the resulting prot related to the corresponding decision branches. In order to maximize total prot, we will choose those branches so that the sum of their prot will be the greatest. In this network representation, the connection of arcs that makes a feasible multi-cycle schedule is identied by two requirements: the ow conservation requirement and the node conservation requirement. Flow is dened as the decision status of an arc. In our model, there are two possible statuses for each arc. We dene 1 to be the status of choosing as part of schedule and 0 to be the status of discarding as part of schedule. As a result, the ow conservation requirement requires that the ow on every arc must be either 1 or zero. The node conservation requirement requires that the sum of ows into every node (inows) must equal the sum of ows out from that node (outows). Consequently, in our problem the net ows (inow minus outow) of node 1 to node T 1 are all zero and there is one outow emerging from node 0 and one inow into node T. A ow path that starts from node 0 and ends at node T and abides the node conservation requirement and ow conservation requirement denes a potential multicycle management schedule. In Fig. 3, the number above the arc denotes the status of ows. Under the two conservation requirements, every node at most has one inow and one outow. Once a node has an inow, it must also have an outow. The feasible multi-cycle schedule presented in Fig. 3 is H((0, i), (i, T k), (T k, T )). The optimal scheduling problem for a single pond is in fact equivalent to determining the ow path that maximizes total prot for that pond. Fig. 4 illustrates how we can compute the resulting prot for a potential harvesting through the underlying biological and economic functions implemented using a set of spreadsheet LOOKUP tables. For each arc, there is a corresponding prot (or loss). Therefore, by multiplying every arcs ow with its prot, we can obtain the prot for a ow path that denes a feasible multi-cycle management schedule. We now consider managing N ponds together and thus we need to incorporate the labor force and market demand constraints. The labor force and market demand constraints are introduced into the model through the node-ow conservation requirement. In the network representation, the arc is counted as an outow to its beginning node, while it is counted as an inow to its ending node. Since the beginning node represents the restocking time and the ending node represents the harvesting time, an outow thus indicates a managerial decision on restocking and an inow indicates a managerial decision on harvesting. In consequence, the labor force and market demand constraints or the restrictions on harvesting and restocking can be imposed upon the sum of inows and outows of each node, respectively. Note that a node can be a beginning node and ending node at the same time. For example, in Fig. 1, node 2 is the

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

347

Fig. 4. Evaluation of prot through the underlying biological and economic functions within the scheduling problem.

beginning node for arc f, while is the ending node for arc b. Hence, for node 2, arc b is its inow, while arc f is its outow. For our present situation, node-ow conservation requirement species that the sum of inows of every node except node 0 across all ponds must abide the minimum and maximum harvesting frequency constraints and the sum of outows of every node except node T across all ponds must abide the minimum and maximum restocking frequency constraints. It is rather straightforward to obtain both the sum of overall inows and the sum of overall outows for every node by directly adding the inow and outow across all ponds respectively and implement the constraints. The solution of optimal scheduling with the labor and market demand constraints is obtained by nding the feasible combination of ow paths for all ponds that produces the highest prot.

4. An application In this section, we apply our optimal scheduling network model to an existing shrimp farming entity to test its applicability, validity and reliability. This commercial shrimp farm under investigation currently operates 40 one-acre ponds throughout the year. Presently, the year ahead harvesting and restocking schedule is decided by a manager through an educated trial-and-error process. So, we will test our present network model based on the same information as used by this manager in order to compare the model schedule and the schedule developed by this manager.

348

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

Fig. 5. The network structure: an application.

Specic assumptions, background information and functional forms for this application are as follows:1 (1) All of the 40 production ponds are homogeneous in terms of environmental and production conditions. The production schedule comprises 52 weekly managerial decisions for a year. Harvesting is assumed to occur only on Mondays, Wednesdays and Thursdays while planting to occur only on Fridays as presently done on this farm. To capture this characteristic, we use subscript a to denote the potential harvesting days (Mondays, Wednesdays and Thursdays) and subscript b to denote the potential restocking day (Fridays) for each week. We dene these two types of node as harvesting node and restocking node respectively. As a result, two types of arcs are allowed in this present network (see Fig. 5). One is the preparation arc, which starts from a harvesting node to another later restocking node and denes the resting time before a new planting occurs. For example, arcs (11a , 14b ) and (31a , 33b ) in Fig. 5 are preparation arcs. The other is the production arc, which starts from a restocking node to another later harvesting node and denes a production cycle. For example, arcs (14b , 31a ) and (33b , 51a ) in Fig. 5 are production arcs. The resting time between two production cycles is restricted to range from 1 to 4 weeks, according to this farms present practice. As a result, the preparation arc connects each harvesting node to each later restocking node that is at least 1 week ahead but no more than 4 weeks ahead. For example, the arc (11a , 14b ) in Fig. 5 implies that the farm keeps the pond empty for 3 weeks before a new production cycle starts. As for the arc that starts at node 0, it will end at a node with subscript b if the pond is empty at initial planning time. Otherwise, it will end at a node with subscript a. Node 53 denotes the ending time. All ponds are assumed to be harvested at that time. Hence, all ow paths will end at node 53. (2) The growth function is assumed to depend only on weight and temperature, because the amount of feeding is assumed to be at optimal levels in order to maximize growth given weight and temperature. This shrimp farm has already generated a weekly growth chart based on its historical sampling data. This chart provides the growth information in this application. This growth function has two essential features that signicantly inuence our optimal managerial schedule. One feature is the seasonality of growth caused by seasonal temperature differences. Incremental growth generally increases from spring to summer, reaches its peak in May, and then decreases gradually until it reaches its trough in December. The other feature is related to the weight dependency in that incremental growth declines once the shrimp grows up to 21 g. (3) Harvest is assumed to occur when the average weight of shrimp falls between 21 and 25 g. This reects the current practice of the market targeted by this operation. This feature is incorporated by assigning zero to production when the weight deviates from the range of 21 to 25 g. (4) Average shrimp weight for planting is assumed to be identical and equals to 2 g.
In order not to reveal any condential information of this commercial shrimp farm, we will be presenting many of the information in a generic form just for the purpose of illustration.
1

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

349

(5) For our present scheduling application, we used a recent historical snapshot from this farms record as the starting point. Hence, the 40 ponds are assumed to be at different points during their production cycles at the initial planning time. In particular, 36 ponds have already been stocked with shrimps and 4 ponds are empty. (6) Average stocking number is assumed to be identical and equals 400,000 per pond. Survival is assumed to be 50% from stocking to harvest size between 21 and 25 g. (7) Prices and costs were both ignored in this present application. Thus we are in essence only maximizing total production and not prot in this application. This simplied assumption is implemented merely for illustrative purpose in testing this model so as not to unnecessarily disclose condential information of this farm. (8) No more than three harvests per week can be performed and there must have at least one harvest per week. There is no specic restriction on restocking. It should be noted while some simplifying assumptions were made in this application for illustrating the working of this model; it nevertheless reects very closely the current practice of this farm according to its vice president.

5. Baseline and alternative simulation results The above network model consists of 30,000 binary variables and is implemented on MS Excel using Frontlines Large-Scale Linear Programming Solver. It takes several hours for Solver to crunch the optimal managerial schedule. Fig. 6 provides the optimal model managerial schedule for the planning horizon of 1 year for the baseline simulation. All harvests are guaranteed to result with shrimp within the target size range by this schedule. In contrast, the existing educated trail-and-error managerial schedule could not provide such a guarantee. The total production under the optimal model schedule is estimated to be 5% higher than that based on the existing schedule. This estimated increase is equivalent to production from an additional two ponds which is not an insignicant amount.2 In order to extract some insights from the optimal scheduling problem, several other scenarios are also simulated, including the scenario where all ponds are empty at the initial planning time and the scenario where no labor force and market constraints are applied. Several interesting ndings can be discerned from comparing the results of these simulations and they are summarized as follows. (1) Even without considering the labor force and market constraints, the optimal harvest size varies depending on the assumption of the length of preparation time. Suppose at the end of time t, the average shrimp weight rstly reaches the target size range, e.g. 21 g, so that we can harvest at the beginning of time t + 1 and start a new cycle after a rest period or keep growing the shrimp to a larger size at time t + 1. The expected benets will be NG21+ if we keep growing the shrimp to a larger size t+1
It should be noted that the management of this operation is very experienced and highly sophisticated. Thus the trial-and-error managerial schedule is certainly not crude but rather a very educated schedule. While we could not reveal the exact dollar amount of the 5% difference, it is sizable and the management has also indicated to us that our present model is certainly worth the effort particularly the model would provide a exible and automatic tool for their periodic scheduling task. This would also imply that operations with less sophisticated management skill would even benet more from the present model.
2

350

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

Fig. 6. The optimal managerial schedule.

while the opportunity cost will be NG21 if we harvest the shrimp at time t + 1 and start a new cycle t+i at time t + i (where N is the surviving shrimp number and assumed to be identical). If the next cycle can be implemented without any preparation time, or i = 1. That is the new cycle starts at time t + 1. In this case, the expected benets will be NG21+ and the opportunity cost will be NG21 , respectively. t+1 t+1 Because the assumed growth function species that smaller shrimps grow faster than larger ones (i.e., G21+ < G21 for all t), then NG21+ < NG21 . Therefore, once the average shrimp weight reaches t t t+1 t+1 the target range (2125 g), we should harvest the shrimp immediately, because the expected benets of harvesting shrimp at a larger size at next time period is always less than the opportunity costs. Since the smallest target weight is 21 g and the average weekly incremental weight is roughly 1 g, we can derive that the optimal harvest size will fall between 21 and 22 g if the next cycle can be implemented without any preparation time. However, due to seasonal variation in growth rate, G21+ maybe greater t+1 than G21 , if the assumed resting period is between 1 and 4 weeks, or the new production must t+i start at least 1 week after and within 5 weeks (i.e., 2 i 5) as in our present problem. In this case, the optimal harvest size may fall into the range of 2225 g. Results of the simulations (see Table 1) indicate that the dominant optimal harvest size is between 21 and 22 g. In fact, 68% of the all 106 harvests fall within this range. The optimal harvest size of the remaining harvests ranges from 22 to 25 g. (2) The optimal harvest sizes of the baseline are found to deviate from those when there are no labor force and market demand constraints. Fig. 7 shows that when there are no labor force and market demand constraints, the optimal harvest schedule is not evenly distributed throughout the year with many weeks of no harvest as well as many weeks with more than three harvests. Obviously, with the imposition of the labor force and market demand constraints, some production cycles would

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354 Table 1 Comparison of managerial schedule with/without the labor force and market constraints With constraints Total number of harvests Average number of weekly harvests Standard deviation of number of weekly harvests Weeks with no harvest Weeks with 13 harvests Weeks with 3+ harvests Maximum number of weekly harvests Average weight at harvest (g) Number of harvests with average weight 2122 g Number of harvests with average weight 2223 g Number of harvests with average weight 2324 g Number of harvests with average weight 2425 g Decrease in total production (%) 99 2.1 0.96 0 52 0 3 23.5 54 16 10 19 5

351

Without constraints 106 2.2 2.81 23 13 16 15 23.2 72 11 11 12

have to be extended and hence with larger harvest sizes than the situation without the constraints. Results of our simulations indicate that overall production would be reduced by an estimated 5% level.

Fig. 7. The optimal managerial schedule without labor and market constraints.

352

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

(3) The labor force and market demand constraints dramatically change the overall optimal management strategies and they are considered the critical factors of the multi-pond and multi-cycle scheduling problem. Table 1 provides a comparison of the managerial schedule with and without these constraints. Without these constraints, the average number of harvests per week is 2.2 and its distribution is greatly uneven. The schedule indicates that there are 23 weeks with no harvest at all and 16 weeks with more than 3 harvests. Especially, in week 36, it has 15 harvests. On the other hand, with the labor force and market demand constraints, the average number of harvests per week is reduced to be 2.1, because fewer harvests are allowed within the year. The derived schedule shows that there are 23 weeks with three harvests and 19 weeks with one harvest. The smoothing of production to meet labor requirement and market demand throughout the year is certainly more realistic and representative of actual shrimp farm operations. However, as mentioned in (2), this comes at a cost of roughly 5% of production reduction. (4) The model managerial schedule is very susceptible to the initial stocking conditions. We may not be able to arrive at a feasible managerial schedule under certain initial stocking conditions. One obvious example is the scenario where all ponds are empty initially. It is easy to see that the condition of at least one harvest per week cannot be satised at all. Similarly, because the target shrimp size is between 21 and 25 g and most weekly growths are greater than 1 g, given the constraint of at most three harvests per week as in our present application, it is very possible to fail to nd a feasible managerial schedule when there are many ponds stocked with similar shrimp size initially. For example, the scenario where 15 ponds have their shrimp sizes between 5 and 6 g initially fails to generate a feasible managerial schedule. Our simulations, using the actual operational data, however, do not encounter this problem. It implies that this commercial shrimp farm actually operates the growout entity quite efciently through trial and error and/or learning by doing. In summary, our model managerial schedule outperforms the farms existing schedule in three respects. First of all, our model managerial schedule guarantees that all harvests will result with shrimp within the target size range. Second, under the model managerial schedule, the harvest shrimp size is essentially determined by weighting between the expected benets and opportunity costs while under the existing schedule, there is no explicit consideration of benet and cost at all. Finally, the model managerial schedule is generated by systematically maximizing the overall production and thus incorporates the interactions between ponds due to the labor and market constraints, while the existing schedule can not systematically synchronize these interactions at all. Hence, it is not surprising that our model managerial schedule can improve the farms performance as our simulations documented.

6. Concluding remarks In this paper, we introduce a practical scheduling model for a shrimp operation practicing multiple cycles per year in a multiple ponds setting to derive the optimal harvesting and restocking schedule. By converting the scheduling problem into a network formulation allows us to solve this otherwise unsolvable hard integer programming problem. The successful application to an existing shrimp operation in Hawaii has illustrated the models ability and reliability in improving the quality of scheduling decisions and the resulting production for shrimp culture. The network formulation of this scheduling problem can be concluded to be a useful tool for supporting management decisions. The present model can be applied

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

353

to other aquaculture operations utilizing multiple production units without changing much of the basic structure since the model is constructed in a very general form as described in Sections 2 and 3. It should be noted that many of the model features have not been utilized in the present application which may be deemed essential in applications in other settings such as the seasonality effects of price and demand on the optimal schedule. Further experimentation with the present model will no doubt provide further insights in improving the present model structure. Besides using the model as an operating tool in deriving the optimal harvesting and restocking schedule, the present model can also be useful as a design tool such as to identify the optimal number of growout ponds given the farms labor and market demand constraints. The present model can also be extended to include the nursery and hatchery phases of shrimp farming so as to synchronize production between these and the growout phase. The present modeling framework certainly provides a solid foundation for these and other possible model extensions for improving the management of aquaculture operations.

Acknowledgements The authors thank Dr. Paul Bienfang of CEATECH USA, Inc. who has contributed signicant insights into the construct of the present model and also for providing valuable farm information without which this study would not be possible. This study was made possible by the generous support of the United States Department of Agriculture Cooperative State Research, Education and Extension Service (USDA grant 2003-34167-14034).

References
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] R. Arnason, Optimal feeding schedules and harvesting time in aquaculture, Mar. Resour. Econ. 7 (1992) 1535. T. Bjorndal, Optimal harvesting of farmed sh, Mar. Resour. Econ. 5 (1988) 139159. O.J. Cacho, H. Kinnucan, U. Hatch, Optimal control of sh growth, Am. Agric. Econ. 73 (1991) 174183. O.I. Forsberg, Optimal stocking and harvesting of size-structured farmed sh: a multi-period linear programming approach, Math. Comp. Simulat. 42 (1996) 299305. O.I. Forsberg, Optimal harvesting of farmed Atlantic salmon at two cohort management strategies and different harvest operation restrictions, Aquacult. Econ. Manage. 3 (1999) 143158. F. Glover, D. Klingman, N.V. Phillips, Network Models in Optimization and their Applications in Practice, John Wiley & Sons, New York, 1992. R.L. Hean, An optimal management model for intensive aquaculture: an application in Atlantic salmon, Aust. J. Agric. Econ. 38 (1994) 3147. T. Heaps, The optimal feeding of farmed sh, Mar. Resour. Econ. 8 (1993) 8999. T. Heaps, Density dependent growth and the culling of farmed sh, Mar. Resour. Econ. 10 (1995) 285298. E. Hochman, P.S. Leung, L.W. Rowland, J.A. Wyban, Optimal scheduling in shrimp mariculture: a stochastic growing inventory problem, Am. J. Agric. Econ. 72 (1990) 382394. L. Karp, A. Sadeh, W.L. Grifn, Cycles in agricultural production: the case of aquaculture, Am. J. Agric. Econ. 68 (1986) 553561. P.S. Leung, Y.C. Shang, Modelling prawn production management system: a dynamic Markov decision approach, Agric. Syst. 29 (1989) 520. P.S. Leung, E. Hochman, L.W. Rowland, J.A. Wyban, Modeling shrimp production and harvesting schedules, Agric. Syst. 32 (1990) 233249.

354

R. Yu, P. Leung / Mathematics and Computers in Simulation 68 (2005) 339354

[14] P.S. Leung, D.J. Lee, E. Hochman, Bioeconomic modeling of shrimp and prawn: a methodological comparison, Aquacult. Models Econ. 4 (1994) 3953. [15] S. Pascoe, P. Wattage, A. Naik, Optimal harvesting strategies: practice versus theory, Aquacult. Econ. Manage. 6 (2002) 295308. [16] Purwanto, Economic optimization in prawn farming, Ph.D. Thesis, James Cook University, 1997. [17] D.H. Spaargaren, Optimal harvest size in shrimp cultures, Crustaceana 72 (1999) 297306. [18] R.R. Springborn, A.J. Jensen, W.Y.B. Chang, C. Engle, Optimum harvest time in aquaculture: an application of economic principles to a Nile Tilapia, Oreochromis niloticus (L.), growth model, Aquacult. Fish. Manage. 23 (1992) 639647. [19] X. Tian, P.S. Leung, D.J. Lee, Size economies and optimal scheduling in shrimp production: results from a computer simulation model, Aquacult. Eng. 22 (2000) 289307. [20] X. Tian, P.S. Leung, E. Hochman, Shrimp growth functions and their economic implications, Aquacult. Eng. 12 (1993) 8196. [21] D.G. Whiting, H.D. Tolley, G.W. Fellingham, An empirical Bayes procedure for adaptive forecasting of shrimp yield, Aquaculture 182 (2000) 215228.

Das könnte Ihnen auch gefallen