Method for portioning of articles

[0001] The invention relates to a method for portioning articles. In particular, it concerns the method of portioning foodstuff based on at least one property of the articles.

Background [0002] There are many industries in which the sorting, allocation or otherwise portioning of articles is important, especially for storing and transporting these articles. This is particularly important in the food industry, where the business of packing, storing and transporting food from the source to the seller or end user is very important. It is highly desirable to portion food products in the most effective way possible, to protect and secure food products, while optimizing the available space for improving the cost-effectiveness of the method.

[0003] When trying to optimize the number and nature of products placed in one or more containers having a specific capacity and target fill levels, many known approaches use algorithms. A particular problem when filling containers with a predetermined mass of a target product, using a limited number of products of known mass and having a limit on the number of containers that can be opened for filling at any given time, is known as the "problem of packing boxes with limited space on the line". As you know, the problem with finding the optimal box packing is NP-hard, which makes it unlikely that there is an effective polynomial algorithm to find the optimal solution to the problem.

[0004] In the food industry, there are several portioning and classification tasks that are carried out primarily manually because known algorithms are not suitable for performing or automating such tasks. For example, fresh fish fillets and poultry breast fillets are commonly packaged in so-called interlaced parcels, which is one example of a packaging method that imposes an additional burden on the portioning solution. Packets are interlaced with typically frozen blocks. The fillets are placed in layers, where the fillets do not touch each other and a plastic film is placed on top of the layer before the next layer of fillets is placed in the package. Such packages usually choose restaurants, because they are very compact and the fillets are not frozen with each other,

[0005] When packaged in packets of interlaced cod fillets, the most common is the presence of three fillets in each layer and the fillets in the first layer are placed as shown in Figure 1, where the tails are placed towards the center of the packet, while the thicker fillet portions are directed to the edge. The figure also shows how the two fillets have a tail portion directed downwards and one up. When the next layer is placed, it is inverted, because the two fillets now have tail parts facing up and one down, but the tails are still located towards the center. In addition to these requirements, packages must have a constant weight and may also have the requirement that there must be a certain number of fillets in the package.

[0006] Two known alternatives exist for full automation of interlaced packaging. One is to make a more complex sorting machine that can selectively rotate the fillets and then the related portioning algorithm does not have to take into account the orientation of the fillets. Another way is to place each, without exception, a fillet in the packing machine with the tail facing the left, while the other fillets have a tail facing the opposite direction. However, this will create a requirement for the portioning algorithm to be able to select fillets. Current methods that are based solely on statistical or combinatorial methods are not at all suited to such a task.

[0007] In the case of interlaced systems as well as other packaging systems, there is a great need in the food industry to sort and pack articles into parcels of predetermined mass. In recent years there has been some progress in the development of sorting and packaging methods for parcels of a fixed weight. Such methods are sometimes called "intelligent portioning", in which items are selected intelligently for different portions (parts) in order to bring them as close as possible to the predefined target mass. With a view to creating so-called portions of a predetermined mass - i.e., portions that have the smallest possible mass above the predefined minimum mass - there are mainly two methods that are used; storage weighing and combination weighing.

According to the principle of accumulative weighing, the products to be allocated to the container or other packaging or storage means are weighed on the dynamic weight and mass are recorded for the purpose of keeping track of the respective product placement in line and the respective masses. The distribution unit then assigns individual products to one or more receiving boxes, as long as the mass collected in a specific box matches the target weight. Nevertheless, accumulative weighing has many drawbacks. Typically, in the methods, when a decision is made, to which of the many recipient boxes the article will be lead, the mass of only a single article is known. This creates a significant disadvantage, because the method can not ensure that the portion in the recipient box will be completed at a certain mass above the target weight, because it can only predict with a certain probability that the method will be able to complete the portion within a given mass. Moreover, it is not very feasible to use this method to create packages with special requirements, as in the case of interlaced packages.

[0009] US 5,998,740 discloses a weighing and portioning technique based on the "sorting" technique, which is a development of storage weighing, in which a number of portioned articles, namely natural food products of heterogeneous, variable mass, are fed via a weighing station and then selectively. feeds into many receiving boxes. The technique involves weighing the finite number of products to be portioned, using the weight distribution of these products to statistically evaluate the best possible division of articles.

[0010] Another known method is described in patent EP01218244B1, in which instead of generating perspective functions for filling one or more boxes based on a single product, the perspectives are based on the so-called Calculation Algorithm.

Smart Tasting (Tally Intelligent Batching Algorithm). The result of the calculation is the total number of possible combinations of products present in the FIFO queue (First In First Out - first in the output) as a function of the portion deficit and the number of products. The method may use the knowledge of the existence of more than one allocated product of known mass, but since it works with the number of possible combinations, as opposed to trying out real product allocations into boxes, it may overlook good allocations. What's more, it is computationally intensive and inefficient and is not suitable for meeting special packaging requirements, for example when packing interlaced packages.

[0011] According to the principle of combined weighing, the combined mass of many assigned articles is known. Typically, there are many weighing hoppers and articles from any of the hoppers can be released to form a portion to fill the container or box. There is therefore random access to weighing hoppers, and the hoppers that have been emptied can be filled selectively. Using combinatorial algorithms, a combination of hoppers is selected that generate the lowest overflow and subsequent products are released from these hoppers and unified to create a portion. Machines with a weighing hopper that use this method of combination weighing are typically in a circular or linear arrangement, as described in U.S. Patent 4,442,910 and 4,821,820.

[0012] There is another set of methods in which the mass of many articles is known, as in combination weighing, but the products are nevertheless sequentially fed into one or more recipient containers or boxes. One example of such a portioning method is found in EP01060033B1. The method uses a portion collection station to temporarily hold one or more products until their mass is suitable to be added to the box to help achieve its final filling level. The main disadvantage of this method is that it does not perform the sorting at the same time as the portioning takes place. Moreover, if there is currently a small number of articles of known mass in the machine, it can not exactly do the portioning task. Furthermore,

[0013] WO-A-01/07324 relates to the generation of a perspective for filling a portion with at least one product having a characteristic property.

[0014] WO-A-2007/083327 relates to apparatus for sorting articles based on at least one property of articles.

[0015] US-A-2005/0137744 relates to a method and apparatus for separating and processing products.

[0016] WO-A-2006/106532 relates to apparatus for sorting and portioning articles, where items sorted into classes are weighed.

Summary [0017] The invention is described in the claims.

[0018] As a method for assignment of a product within a product group to a selected container from a group of containers is provided, which considers the characteristic property of both this product and another selected product within this group, together with information on the container capacity, can be achieved optimal allocation.

The allocation is not based on predicted or predicted information, nor on statistical trends, but on the actual characteristics of the product and product being allocated, which may be assigned next. The characteristic properties of the products may include any of size, weight, length and orientation or any other property depending on the type of product under consideration. Because the method considers possible options for assigning both the first product and the second product to the group of available containers, one can look at the impact effects of the possible allocation choices for the first product, in particular the allocation of the first product can be chosen so that the feasible options remain for the second product allocation after it. . Therefore, optimization is not immediately beneficial and accurate,

[0019] Since the subsequent addition of many products to a group of containers can be considered, a collective picture of the effect of each allocation option is provided. It increases the intelligence of this method and helps to improve and further optimize the allocation of products. In addition, by considering the subsequent assignment of multiple products to two or more different combinations of containers and comparing the respective impact of these allocation options on capacity information for a group of containers, the effects of available allocation options in the real world are considered and the most favorable allocation option can be chosen for determining to which container the first product should be allocated. Thus, a more accurate solution is provided that can deal with real-world phenomena,

[0020] By making it possible to account for the actual capacity of the container after the product has been allocated to it and / or the expected capacity of the container after the product has been allocated to it, the prognostic possibilities of the method are increased. That is, the method may use actual information about the characteristic property, regarding the products in the groups to extrapolate and predict future capacity effects that may affect the selection of the container to allocate the first product to it. Because the expected capacity of the container after the product is allocated to it is based not only on historical allocation information, but also on current allocation information, in this information about the characteristic property for actually remaining products to be allocated and / or capacity information for actually available containers for allocating these products to them, the prediction can be better refined compared to prior art methods which are based exclusively on previous trends or distributions. This means that when calculating how the capacity of a particular container can change after the product has been allocated to it, any anomalies or non-standard features of actual products or containers can be taken into account.

[0021] By considering both the current fill level and the target fill level for the containers, it is possible to better focus on the product assignment allowing it to be allocated to the container in which it would be most useful in allowing effective achievement of the target fill level and avoiding or at least reducing the loss and potential loss of profits related to the overfilled container. By predicting what the fill level would have been obtained, if a product were added to a particular container within the available group of containers, a clear picture of how best to allocate the product or other product in the group can be achieved.

[0022] In particular, since the obtained filling level of the container can be considered if any number of different containers within the available group of containers is added to another, the second product can be forecasted and planned as part of the allocation method. This means that an allocation that would work well for the first product could turn out to be less than optimal when considering the second product.

[0023] By going one step further and looking at the level of filling received, if one or more available containers were allocated to each of the products in the product group available for allocation, optimization and future planning could be increased. This benefit is not only obtained for the allocation of the current product group, but it can also be used when planning and controlling the sorting and provision of subsequent product groups for allocation. Nevertheless, even if the possible allocation of each product in the product group is considered, only the allocation of the first of these products will be determined as a result. Ideally, whenever a new product within a product group is to be allocated by determination

[0024] By limiting the time at which assignment determination is performed, a balance is achieved between forecasting to improve the optimization of the allocation of a single product and maintaining the course of the allocation method at an acceptable rate.

[0025] By failing to consider adding a product to a particular container or combination of containers when it appears that such assignment option provides a less favorable result than the result which, as shown, has just provided an alternative option, the computational efficiency and speed of the allocation method is improved. Moreover, prioritizing a particular container or combination of containers based on capacity information enables faster and more efficient achievement of an optimal allocation solution. Likewise, ignoring a container or a combination of containers according to capacity information avoids wasting time on "dead end" options, thus improving overall efficiency and speed.

[0026] Flexibility and additional control are provided by allowing user-defined constraints. Depending on specific product assignments and other real-world conditions, user-defined constraints can include any number of factors, including the number of products in a product group, the number of containers in the container group, many container allocation products, and, at any time, the limit for filling the container to the target level before replacing it with a new, empty container. Furthermore, the orientation of the product being allocated and its configuration with one or more other products after being allocated to the container may be considered. Thus, special packaging systems, such as interleaving, can be effectively and usefully incorporated in the method of attachment.

After determining the allocation, the first product may be directed to the selected container.

[0028] By predicting future volume information for a group of containers, using the allocation determined for the first product and the characteristic property of at least one of the other products in the product group, the method enables intelligent selection of subsequent products for assignment and assisting in determining the allocation for this subsequent product. In particular, since the future level of filling of the container can be predicted, the allocation of the current group can be adjusted accordingly and / or the future product group can be intelligently selected to be allocated to containers for receiving such products.

[0029] By providing information about the characteristic properties and capacity information for the next product being allocated, it is ensured that the most appropriate and the most accurate information available is used all the time rather than relying on previously obtained or assumed trends or patterns that could be used to earlier products, but for which it is not possible to guarantee equally good use for subsequent assigned products.

[0030] Accordingly, a method and associated control and operation are provided that have significant advantages over the prior art methods. The approach is accurate, effective, intelligent and flexible, while being affordable to implement with existing transmitting and receiving devices and for different types of products.

The embodiments will now be described with reference to the figures, among which:

[0031]

Figure 1 shows first and second interlaced layers of packaged articles, such as fish fillets;

Figure 2 illustrates an exemplary machine layout;

Figure 3 shows a prospective search at a depth of 1 product and 2 products;

Figure 4 shows the expected loss from a single box as a function of its capacity; and

Figure 5 shows the expected total loss from many boxes as a function of their capacity.

Review [0032] The method relates to portioning articles of known size, also known as "classification", the classification comprising gathering portions of articles in containers or classifying articles on one or more recipient areas. The method uses the search method to find the optimal receiving container or receiving area in which the article is to be placed, based on one or more of the many possible classification criteria. The most common feature taken into account for classification purposes is the mass of articles and the most common criterion is to prepare a portion as close as possible to the previously defined desired portion weight.

[0033] The method mainly concerns the problem of packaging food products, e.g. fish fillets, into containers, so that the total overweight of filled containers is minimized. The container is treated as full when its mass reaches or exceeds the predefined target mass, when overweight (or loss) is defined as the difference between the final weight of the container and its target mass. In one possible configuration, dictated by grading and portioning machines on the production line, the maximum number of simultaneously opened containers to which the fillet can be allocated is determined with the current actual weight of each container and container replaced with a blank piece after filling (possibly with a certain delay) ). Moreover, at any given time, the size of 1 or more subsequent fillets to the vessel is known.

[0034] The problem here can be formulated as a problem of packaging boxes with limited space on the line, as briefly mentioned in the background section of the invention above. More specifically, the task is to allocate products X = {x1, x2, ..., xn} of size (0.1) to the unit-size boxes, to maximize the number of filled boxes to a size of at least 1. The task is a limited space K, because the number of open boxes can not exceed K at any time during the work.Moreover, the task is in real time (online - on the line), because the products come one at a time and have to be allocated to the box after the arrival. the examined problem of packing boxes, where only the size of the product allocated (xi) is known, it is assumed that the size of the upcoming "m" products is known, including the one to be currently assigned. That is, when product xi is allocated, the size of products xi, xi + 1, ..., xi + m-1 is known.

[0035] As described in detail below, an effective approximation algorithm is provided for carrying out the allocation of articles, which makes good use of all available information. Moreover, it allows for greater flexibility in imposing various restrictions on reasonable allocations, e.g. in relation to a combination of products placed in a box (e.g. based on the orientation that the products can be arranged in layers), or the speed at which boxes are being replaced.

[0036] The method described herein may work with one or more additional requirements imposed by a user in collecting a portion, such as selecting only articles of a certain orientation, which may then subsequently impose a requirement on the orientation of future articles placed in the respective portion.

DETAILED DESCRIPTION OF THE INVENTION [0037] In the application, for coherence with the available literature on the subject of packaging boxes and to emphasize the overall applicability of the proposed algorithm, the following terminology was adopted:

Pieces allocated to containers are referred to as products, and containers are referred to as boxes. The most important property of a product is its size, which determines how much storage space it occupies. The size may include one or more features, including mass, length, volume or density of the product. In the description below, the main size characteristic used is mass.

[0038] The storage capacity of the box is called the demand, and the box level is the sum of the size of all products in the box. If the box level is equal or exceeds its demand, we say that the box is closed. Otherwise it is open. The loss is defined for a closed box and it is the difference between its level and its demand. For example, in fish file packaging applications, the box corresponds to the container, the fish fillet product, and the product size to the fillet weight.

[0039] Compared with the more traditional packaging problem of boxes, where only the size of a single incoming product is known due to its allocation to a container, it is known that, according to current approaches. A significant influence on the design of an effective algorithmic solution is knowledge of the size of the incoming products in the input sequence. The use of an existing algorithm for the traditional packaging problem would lead to not necessarily suboptimal solutions because important information is ignored. Such algorithms, in which only the size of a single product currently allocated is known, must base their decisions on the allocation only on the expected size of future incoming products. For example, if products of small size are rare, the prior art algorithms would avoid leaving the box in a position close to its requirements. Nevertheless, if there was knowledge of such a small element that would come soon, leaving the box level in this position would be a perfectly sensible thing to do and could lead to a better solution than otherwise possible. This type of knowledge and control is available according to current approaches and therefore, as a result, improved allocations are possible.

[0040] For example, consider a case where another three incoming products (e.g.

fish fillets) are respectively 400, 600 and 100, and there are two boxes partially filled up to 4400 and 4500, respectively. What's more, let's assume that the goal is to fill boxes with a demand of 5,000 and that the average product size is about 400. A way that looks at only the incoming products and distributes them to the boxes based solely on statistics is required to place the first product in the first the two boxes mentioned. The reason for this is that the box with a level of 4800 has better room for the next product than the one from 4900, based on the known and assumed mass distribution of the current group of products to be allocated. However, this prior art approach leads to a poor solution. On the other hand, using the current approaches and looking at all of the actual three products when performing the separation, there is an optimal solution for placing the first and third products in the second box and the second product in the first, which leads to no overweight in this case. Therefore, an efficient algorithm, when allocating product xi to the box, should also take into account the size of products xi + 1, ..., xi + m-1.

Search [0041] As will be appreciated by the person skilled in the art, at least in theory, for the development of all different possible m m products to possible boxes, the force search can be carried out and then it evaluates the value of each possibility and selects the most promising one. However, there are practical problems with such a strength approach. First, because the number of possible allocations using such a force approach increases exponentially in m, at a rate of 0 (K^m), it is impractical to use all but the smallest values ​​for K and m. For example, even using a rather modest statement K = 4 and im = 12 leads to more than 16 million possible combinations. Secondly, you need to have a quick calculation mechanism to predict the future expected loss for partially filled boxes, as this assessment is required to evaluate each of the many possible allocations. Assuming that every opportunity can be found and evaluated in only one millisecond, the decision to which box to send xi would still take over 4 hours. It is purely impractical.

[0042] According to current approaches, the search algorithm based on prediction at any time is used for this task. Preferably, the algorithm uses one or more techniques, including cutoff bounds, action ordering, progressive lookahead, and an effective predictor of a future loss (future-predictor) .

The algorithm with these improvements thus reduces the search space by several orders of magnitude, thus fulfilling the imposed real time constraints that occur in practice, still returning high quality assignments with a small loss.

Allocation [0043] The search algorithm based on prediction can be used at any time to allocate many products to the appropriate multiple containers.

[0044] A preferred assignment method, LOOKAHEAD_ASSIGN (X, B), is described as Algorithm 1 and 2 below. The former is a controller of any time that initiates the assignment method and calls the other to apply progressively deeper predictions, the latter performing the actual prediction search (DFS_BnB) which reveals various possible product assignments to the boxes. Variables X and B are arrays containing, respectively, the size of known products (xi, xi + 1, ..., xi + m-1) and the level of boxes. Table A contains the assignment currently tested in the forecast search, and Amin stores the best allocation found so far. The allocation contained in the table A at any given time may be partial, i.e. until now only some products have been allocated, or may be total,

Algorithm 1 LOOKAHEAD-ASSIGN (Χ [1, Β {]) £ Amuf j * - 0

2: for ra = 1 to length (X) to

3: U> min «- OO

4: Xf] ^ 0

5: DF $ JBnB (m, Q, A)

6: if (time-isjup) then

7: break

8: end tf

9: end for

10: return Λ_ηηη[1]

Algorithm 2 DFSJSnB (i, w_acluoand, A [])

Global: X [], B [], A mt »i (] and TUmin

1: if (lOaeiuai - W min) then

2: return

3: else if (i> length (X)} then

4: Wtotai * - Wac-tuai + expected (B)

5: if (in (pt_and/ 'dt w_miri) then,

6: - ^ rnin *

7: lUniin * W (_ABOUTŁai.

8: end if

9: return

10: end if

11: order (J3, O)

12: for o = 1 to length (O) to

13: b = O (o]

14: if (aZ / o? Ned (i, ó, #, A)) then

15: Λ (ί] «- b

16: in₀ + - assign (X (ij, £ [i>])

17; DFS-BnB (i + 1, Wactual + Woi 4)

18: deasstpn (X [i], B [b])

19: if (timejs-up) then

20: break

21: endif

22: end if

23: end for [0045] Algorithm 1 starts with the initialization of the best allocation as empty. That is, all products are initially unallocated. Then it can trigger Algorithm 2 for progressive implementation of an increasingly deeper prognostic search (rows 2-9) to evaluate what the optimal allocation of the current product is, bearing in mind the next (m-1) products. The variable m tells algorithm 1 how far to look forward. That is, when m = 1, the DFS_BnB procedure only looks at the product X [1] (xi), when m = 2, it looks at products X [1] (xi) and X [2] (xi + 1) etc. This continues until all products are included or the time allowed expires (rows 67). This means that the algorithm can return the solution at any time.

[0046] It can be seen that, although a trial total allocation of all products to X is made, the algorithm ultimately engages in the allocation of the first product (row 10). This is because, at a later stage, the size of the new product will become known in the allocation method as part of the product group to be included for the allocation, which leads to a completely different optimal allocation.

[0047] As will be further understood with reference to Figure 3 below, the DFS_BnB procedure according to Algorithm 2 travels the search tree in a left-to-right manner, first depth, if necessary returning to its stages. The search tree branches are defined by various possible combinations of products to be allocated to available boxes. The function takes two arguments: the index of the product being allocated (range

1 to m) and the sum of the actual loss in boxes, which is covered by the current (partial) allocation.

[0048] Two conditions cause the algorithm to return or continue to look for a better alternative assignment. First, if the total allocation (rows 39) has been made, so that all products m, which are perceived at the current depth as being allocated to the sample, are retained, the algorithm returns after evaluating the received box condition for the expected loss. The "expected loss" approach is further discussed below. If the new best assignment was found as part of such a total (trial) allocation method, it is stored in Amin (row 6).

[0049] Second, the algorithm returns during the allocation method, when the loss already accumulated on the branch in question exceeds the quality (i.e., the magnitude of the loss obtained) of the best allocation found so far (rows 1-2). This return point is known as the "cut-off limit" and is important for early cut-off of non-executable partial allocations. This so-called cutting strategy based on the branch and the boundary is the more effective the sooner a reasonable good solution is found with a small loss. Thus, in order to maximize its effectiveness, the method preferably re-orders the possible assignment activities, so that it is more likely to discover the good than the worse earlier.

[0050] The order function (line 11) decides the order in which open boxes should be included, as well as discard boxes that are probably unsuitable for product xi. In particular, this is to arrange the boxes so that at the next iteration an Amin box [i] is first tested, if available. In lines 12-23 successively different ways of allocating product xi to open boxes are tested. The assign function places the product xi in the bi box. If the box is packed, i.e. it has reached its target filling level and is thus full, the function returns the loss for this box and empties it. When the box is not packed, the function returns zero "0". The deassign function undoes all changes made by assign. The permission function checks for user-defined restrictions, whether the product xi has permission to enter the box b; if not, this action is not taken into account.

Calculation of the expected loss [0051] In the above-mentioned search and assignment procedure, with each leaf where all products m were allocated to the boxes, i.e. at the end of the branch for a given product depth, the overall quality of the specific allocation is evaluated (line 4 in algorithm 2) . The assessment is based on two elements: the first, the actual (ie received) loss for boxes that were filled to their target levels during the allocation of m products (if any), accumulated in the Wactual variable, and the second, expected future loss of hollow and partially filled boxes remaining in the system (calculated by the expected (B) function). Due to the frequency of such calculations, i.e. with each leaf, it is important that they are computationally effective.

[0052] Algorithm 3 below provides details of the expected calculations of a future loss, where the function f is that provided in Figs. 4 and 5.

Algorithm 3 expected (B [J] ___ · _

Global: niean.

SUtotal * 0

2: for 6 = 1 to length (B) to

3- ♦ - capacity-ofJnn (b} - 2? {6j

4: Utotal * Ut_abouttal 4 * tifcoiffej '

5: end for

6: V)_CXP «- 0

7: for 6 = 1 tolength (B} to

8: ^atremaining * ^ totai ⁱⁿhoi [6]

9 '. IN_eXp <- W_esp + / (titox {6j, [UreTnatntng / tTiećTlj) 10: end for

11: return in_{ea; and},.

[0053] Algorithm 3 calculates and returns the expected loss of partially filled boxes. It begins by calculating the total unused box capacity accumulated for all available boxes (rows 1-5). Then for each of the boxes (rows 7-10) their expected future loss is calculated as a function of their unused capacity and unused capacity of the remaining boxes, as shown in calling the "f" function in line 9. The first argument of function f is the unused capacity of the box itself, and the second argument is the unused capacity of the remaining boxes normalized as the number of products expected to match the remaining capacity of the boxes. This standardization is done by dividing the unused capacity by the average size of the product for the m products allocated. This second argument determines which expected future loss curve to use (see, e.g., Fig. 5). At the end, the total expected future loss accumulated for all boxes is returned (line 11).

[0054] As mentioned above, the expected loss of a single box is a function of its unused capacity as well as the size distribution of the packaged products. Fig. 4 shows an example of such a relationship, assuming, as we do, that the size of the packaged products has a normal distribution - in this particular example the size distribution of the product is N (500, 55). The X-axis shows the current unused box capacity in grams, and the Y-axis shows the expected loss in grams.

[0055] The graph in Figure 4 tells us about the expected loss for a box after it has been filled as a function of its current unused capacity, knowing that the incoming products are from the aforementioned distribution; for example, if the currently unused box capacity is 1 kg, we can expect a final loss of about 150 g. This example assumes that the box is the only one available. Nevertheless, if there are other also available boxes that are only partially filled, we have more degrees of freedom in choosing which box to send the product to, which generally results in a total less loss. Thus, when assessing each individual box, the total unused capacity of the remaining boxes must also be considered. For the sake of simplicity and abstraction, we only follow the total unused capacity of the other boxes. Fig. 5 shows the various expected loss functions based on a number of different total unused capacities - the higher the capacity, the less expected loss.

[0056] Thus, current approaches may not only determine the actual loss that could arise for containers that are filled during a specific allocation of m products, but may also go a step further and calculate the expected loss that will arise after filling the remaining available boxes to their levels target. Such calculations are very important in assessing the relative value of the various allocation options for placing the group of m products in a plurality of containers, since the overall purpose of the assignment is to effectively fill the boxes with the smallest possible loss.

[0057] Calculations of the expected loss according to current approaches are advantageous compared to the prior art methods because they use both the actual remaining capacities in available boxes or other filling containers as well as actual size information for individual products in the group of products to be allocated, when returning expected loss value. Therefore, they do not rely solely on the trends observed for previously available products or containers when predicting future losses, but also base the prediction on real products and containers that will be combined and lead to a loss. Calculations of the expected loss can be returned for each product in the group of m products and for each possible allocation option for this product.

Example [0058] An illustrative example is shown in Figure 3. There are two boxes, each with a requirement of 100, initially in the state [10, 40], and two known products X = {75, 15}.

The skilled person will appreciate that reasonable assumptions can be included in the calculation to improve the efficiency of the assignment while maintaining accuracy. In this particular example, we assume that the expected loss calculation function estimates that the expected final loss of each open box that is currently empty is +5, the expected final loss of each open box that is currently non-empty, but less than half full is +10 and the expected final loss of all other open boxes is +15.

[0059] As can be seen in the upper level of the tree in the figure, a forecast with a depth of 1 product is realized. There are two possible assignment: the first product can enter any box. The algorithm first tries to place the product in the first box, which leads to the condition of the box [85, 40]. There is no real loss here, but bearing in mind the previously mentioned assumptions that the expected loss of the received box condition is +25, for such allocation the total loss Wtotal = 0 + 25 = +25.

[0060] The second possibility of placing the first product in the second box leads to the state [10, 0] (closed box with a loss of +15 and replaced with an empty box), which gives Wtotal = 15 + 5 = +20 and is thus advantageous. That is, a forecast search with a depth of one product suggests placing the first product, size 75, in the second box.

[0061] In the next iteration, a forecast search with a depth of 2 products is carried out, presented as a tree of the lower level. Assuming that we do not change the sequence of actions, the method would try again to place the product first in the box (left branch), then the second product in the first box, leading to the box [0.40], which gives Wtotal = 0 + 10 = + 10 (the first box closed without loss). Because this is the best allocation found so far for this iteration, it is stored. The algorithm has made a complete assignment of both products at this level with 2-product depths, so now it comes back and tries to assign the second product to the second box, leading to the state [85, 55], which gives Wtotal = 0 + 30 = +30. The assignment is inferior to the already found best allocation in this iteration, so the algorithm ignores it and, because it is the last action, goes back one level up and now re-allocates the product to the second box (right branch). However, this closes the second box with an actual loss equal to +15, and because such a partial allocation is no longer better than the best allocation found so far (+10), this branch can be cut off. Again, because the second action is the last at this level and no further return is possible, the iteration ends, the best allocation based on the 2-point depth search is therefore to place both products in the first box. goes back one level up and now re-allocates the product to the second box (right branch). However, this closes the second box with an actual loss equal to +15, and because such a partial allocation is no longer better than the best allocation found so far (+10), this branch can be cut off. Again, because the second action is the last at this level and no further return is possible, the iteration ends, the best allocation based on the 2-point depth search is therefore to place both products in the first box. goes back one level up and now re-allocates the product to the second box (right branch). However, this closes the second box with an actual loss equal to +15, and because such a partial allocation is no longer better than the best allocation found so far (+10), this branch can be cut off. Again, because the second action is the last at this level and no further return is possible, the iteration ends, the best allocation based on the 2-point depth search is therefore to place both products in the first box.

[0062] Thus, as illustrated by the example above, an efficient and effective method is provided for evaluating the possible product assignment, bearing in mind the properties of subsequent products, while avoiding considering each single combination of available products and boxes. Although a simple example with two products is given, the skilled person will recognize that the principle applies equally to more products and any number of boxes. Of course, in practice, a limited number of boxes will be open for filling at any time. Moreover, the depth of the search tree and the number of m products to be included will depend on many factors, including the speed at which the computer or program that implements the method can process the steps and return the allocation information.

Variations and applications [0063] Any deployed computer or other control means may be used to implement the approaches. Furthermore, any suitable transfer and distribution apparatus may be used.

[0064] An apparatus that is particularly well suited for the methods described herein is described in Icelandic Patent No. 2320. This apparatus, for example, is well-suited for packaging the interlaced packs discussed above. The methods are not limited to working with such a camera. Other types of apparatus may be used, e.g. standard sorting equipment, as described in U.S. Patent No. 5,998,740. Other alternatives are, for example, slotted conveyors. In fact, any type of equipment can be used for this purpose, which facilitates the sequential selection of products into portions based on one or more features.

[0065] According to current approaches, any suitable weighing means may be used to weigh the products to be allocated and to register their masses. The skilled person will be familiar with such an appropriate dynamic weighing means and, for example, microprocessors for mass registration and location of the products to be allocated.

[0066] Instead of or in addition to filling the recipient box to a target approach mass level, it may be used to fill one or more boxes with a predetermined number or range of product numbers. Moreover, instead of a single target level, the box may have many advantageous target levels within a certain range, with some target levels being more advantageous than others. In addition, the filling of one or more recipient boxes can be carried out randomly, without using the branch strategy and the boundary described above, until the level of filling of this box reaches a certain limit below its target filling level.

[0067] The approaches are described above with reference to sequential access to the product, for example, using a conveyor. Nevertheless, random access to products would also work, although random access is not required for the correct implementation of the approach.

[0068] In addition to product allocation to one of the plurality of receiving boxes, the approaches may include the option of rejecting the product. For example, if the available possible allocations of the current product, based on the currently considered products, can not lead to a loss level that is below a predetermined threshold, the product can instead be discarded or deferred for subsequent allocations if the resulting loss would be less than threshold level.

[0069] According to current approaches, any suitable length or number of conveying means may be used. Moreover, any number of boxes or other receiving areas can also be used. Current approaches can determine the limit of the number of points that are available to fill at a given time and can, as a result, prioritize filling one or more specific boxes at any given time to maintain the speed of sending full boxes at the required level while working regarding restrictions on how many boxes can be opened at any given time.

[0070] While the approaches discussed above mainly relate to portioning of food, it is noteworthy that the approaches apply equally well to portioning products of any type. According to the type of product and user preferences, appropriate restrictions can be imposed on the allocation method.