WO2003042793A2 - Agent utilisant un modele predictif detaille - Google Patents

Agent utilisant un modele predictif detaille Download PDF

Info

Publication number
WO2003042793A2
WO2003042793A2 PCT/US2002/036813 US0236813W WO03042793A2 WO 2003042793 A2 WO2003042793 A2 WO 2003042793A2 US 0236813 W US0236813 W US 0236813W WO 03042793 A2 WO03042793 A2 WO 03042793A2
Authority
WO
WIPO (PCT)
Prior art keywords
data
inventory
computer program
program product
model
Prior art date
Application number
PCT/US2002/036813
Other languages
English (en)
Other versions
WO2003042793A3 (fr
Inventor
Alexander Renz
Ye Chen
Original Assignee
Sap Aktiengesellschaft
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from US10/208,186 external-priority patent/US7222786B2/en
Priority claimed from US10/281,074 external-priority patent/US7644863B2/en
Application filed by Sap Aktiengesellschaft filed Critical Sap Aktiengesellschaft
Priority to AU2002346417A priority Critical patent/AU2002346417A1/en
Publication of WO2003042793A2 publication Critical patent/WO2003042793A2/fr
Publication of WO2003042793A3 publication Critical patent/WO2003042793A3/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/08—Logistics, e.g. warehousing, loading or distribution; Inventory or stock management
    • G06Q10/087—Inventory or stock management, e.g. order filling, procurement or balancing against orders

Definitions

  • This invention relates to agents. Businesses need to adapt to many competitive pressures. For example, financial markets are increasingly demanding that companies use capital more efficiently; other businesses are seeking global playing fields to maintain growth and diversify risk; customers are demanding service as markets of one and forcing companies toward mass customization; and innovation cycles are continually accelerating.
  • a supply chain is a network of facilities and distribution options that performs the functions of procuring materials, transforming the materials into semi-finished and finished products, and distributing the finished products to customers.
  • Supply chain management (“SCM”) is a business policy that aims to improve all activities along the supply chain. SCM results in improved integration and visibility within individual companies, as well as flexibility across supply and demand environments. As a result, a company's competitive position is greatly enhanced by building supply networks that are more responsive than the current sequential supply chains.
  • SAP AG and SAP America, Inc. provide SCM solutions for product manufacturers to help them reach their goals. Some of the SCM solutions are based on the mySAP.com e-business platform.
  • SAP R/3 component that provides enterprise resource planning functionality.
  • the SAP R/3 product includes a Web Application Server ("Web AS"), an R/3 core, and various R/3 extensions.
  • Web AS Web Application Server
  • R/3 core Web Application Server
  • R/3 extensions various R/3 extensions.
  • the SCM Extensions of R/3 provide various planning, coordination, execution, and optimization solutions that are associated with a supply chain.
  • this invention provides methods and apparatus, including computer program products, implementing and using techniques for monitoring and predicting various data related to an inventory.
  • a data processing apparatus interacts with one or more business computers to obtain obtain one of replenishment and consumption activity data related to an inventory.
  • a model for variations in the inventory is created. The model is used to determine estimates of variation for one of currently planned replenishment activities and currently planned consumption activities. The estimates of variation are used to predict future inventory levels.
  • Creating a model can include applying machine learning to create the model.
  • Using the model for variations in inventory can include using an algorithm including a decision tree.
  • Using the algorithm including a decision tree can include using a classification and regression tree.
  • the classification and regression tree can be constructed using a growing stage and a pruning stage to provide one or more splitting points.
  • the growing stage can include instructions for recursively splitting data using a criterion for maximizing the information resulting from each of the one or more splitting points.
  • the pruning stage can include instructions for removing splitting points.
  • the criterion for determining splitting points can include applying analysis of variance techniques.
  • the size of the decision tree can be selected based on a process including instructions for determining a size of the data; splitting data into multiple substantially equal-sized groups; constructing multiple test decision trees, where each test decision tree includes all of the groups except for one of the groups, such that test decision trees are created in which each group is excluded once; pruning each test decision tree at one or more degrees; testing each of the test decision trees by evaluating each test decision tree with the data excluded from creating that test decision tree; selecting a degree of pruning based on a maximization of the average performance for the excluded data; and pruning the decision tree to the selected degree of pruning.
  • the model can include one or more of data relating to: an update number specifying how many times an activity has been updated, a creation date specifying when an activity was first planned, a current date specifying the date when information in a record was extracted from the activity, an expected date specifying a date at which the activity is expected to be fulfilled, an expected quantity specifying an amount of material required by the activity, source of an order specifying an entity making the order, demand at a source in a temporal window prior to the expected date, actual delivered quantity, actual delivery date, time spread for use with activities that are completed in more than one piece, and activity cancelled information relating to cancellation of an activity.
  • a model to for predicting variability on actual delivery quantity can be created.
  • the model can include a model for predicting variability of an actual delivery date.
  • the model can include a model for predicting an order cancellation probability. It can be determined if a qualitative shift in behavior has occurred, and if a qualitative shift in behavior has occurred data prior to the qualitative shift in behavior can be discarded and a decision tree can be created using data generated after the qualitative shift in behavior.
  • the data can be processed based on the quantity of data. Processing the data can include sampling the data if the quantity of the data exceeds a threshold value. Processing the data can include using additional data if the quantity of the data is below a threshold value.
  • Predicted future inventory levels can be displayed.
  • the predicted future inventory levels can be displayed as a bicumulative display.
  • the predicted future inventory levels can be displayed as a relative density display.
  • An early warning agent can be included, the early warning agent having instructions to receive the predicted future inventory levels and use the predicted future inventory levels to determine whether to order additional stock for the inventory
  • the data can be generated using a radio-frequency identification device.
  • the data can be generated upon the sale of stock in an inventory.
  • the data can be generated upon the transfer of stock in an inventory.
  • An implementation can include one or more agents in a supply chain management system.
  • this invention provides a system comprising one or more agent computers, the agent computers.
  • the agent computers include instructions to interact with one or more business computers to obtain obtain one of replenishment and consumption activity data related to an inventory; create a model for variations in the inventory; use the model to determine estimates of variation for one of currently planned replenishment activities and currently planned consumption activities. The estimates of variation are used to predict future inventory levels.
  • FIG. 1 is a plan view of an inventory early warning agent application to a shelf level monitoring system.
  • FIG. 2 is a plan view of an agent framework architecture that includes three agents.
  • FIG. 3 is a plan view of a directory service of an agent community that includes three agent frameworks.
  • FIG. 4 is a plan view of the messaging architecture of the agent community of FIG. 3.
  • FIG. 5 is a plan view of a communication sequence between a pair of agents.
  • FIG. 6 is a graph illustrating cumulative inventory variables over time.
  • FIG. 7 is a flow chart illustrating the operation of an IEWAto predict an inventory level.
  • FIG. 8A is a chart showing real-time demand vs. orders with a three-day response time.
  • FIG. 8B is a chart showing real-time demand vs. orders with a ten-day response time.
  • FIG 8C is a chart showing an auto-correlation function for a time-series of orders.
  • An agent is an active program that performs some high-level business functions, such as monitoring an inventory for a particular Stock Keeping Unit (SKU) for potential stock out situations and sending alerts when various thresholds are reached.
  • SKU Stock Keeping Unit
  • an agent can be programmed as an inventory early warning agent ("IEWA") that monitors one or more inventories in a factory, warehouse, store, or shelf within a store.
  • the IEWA is programmed to monitor the inventory based on real time data that is obtained, for example, from cash registers or radio-frequency identification devices that scan information relating to an item being purchased or transferred and use that information to update the IEWA as to the level of stock on a particular shelf.
  • the IEWA is a predictive and adaptive inventory management application that can be used to monitor and predict future inventory levels by modeling variability in both demand and supply related supply chain activities.
  • the IEWA uses learning techniques that can estimate potential variation in inventory levels in the near future in order to identify potentially risky situations early enough to allow for corrective measures.
  • the IEWA is operated with minimal human intervention, parameter adjustment, and manual calibration. These characteristics are achieved with machine-learning techniques to recognize patterns of behavior from historical data around consumption and replenishment and around the performance of resources and supply chain partners.
  • the IEWA can use two types of modes of operation for its predictive function: in a first mode, the detailed mode, use the IEWA predictive models built from detailed activities; and in a second mode, the aggregate mode, use the IEWA predictive models built from the aggregation of multiple activities.
  • potential variation in inventory is estimated for each planned replenishment or consumption activity. For example, given historical performance data, the IEWA estimates whether a planned truckload of 12 ounce bottles of Bob's bottles will arrive any time within four hours prior to and six hours after the planned delivery time with a quantity that is between 95% and 100% of the requested quantity. A second stage of processing then combines the estimates of potential variation for individual activities into an estimate of the potential variation of the projected inventory levels.
  • potential variation of the projected inventory level is estimated directly from projected cumulative replenishment and consumption. This estimate is calculated as the aggregation of replenishment and consumption activities.
  • the detailed mode of operation is advantageous when low volumes of detailed data are available and when the degree of variation in activities is highly dependent on the particular resource or partner involved.
  • the aggregate mode of operation is advantageous when it is necessary to deal with high volumes of data, or when it is difficult to track the outcome of any particular planned activity.
  • the aggregate model is discussed below initially followed by a discussion of the detailed model. Nonetheless, the examples and descriptions of FIGs. 1- 5 are equally applicable to either mode of operation.
  • a store 100 can want to have a minimum quantity of items 105 on a shelf 110 at any particular time such that the shelf can be stocked with as large a variety of items as possible. For example, by limiting the quantity of a particular brand of tomato soup on a shelf at any particular time, a grocery store can display more brands of tomato soup on that shelf. To accomplish this type of shelving scenario, the store must carefully monitor the amount of stock on the shelf on a real time basis to prevent the shelf from emptying out. Currently, many stores use their employees to visually monitor empty space on the shelf and replenish the shelf when it looks empty. This is labor intensive and inexact. As illustrated in FIG.
  • the IEWA 115 receives shelf quantity data from a shelf-monitoring program 120 that, for example, monitors the quantity of a particular SKU number on the shelf.
  • a shelf-monitoring program 120 that, for example, monitors the quantity of a particular SKU number on the shelf.
  • the IEWA 115 uses algorithms to determine whether there is likely to be an undesirable variation in inventory and, if there will be one, when to send an order to replenish the shelf 110 with items of that particular SKU.
  • the IEWA 115 sends a message to a planning module 125 along a first message path 130.
  • the planning module 125 then sends return messages to the IEWA 115 along the first message path 130 and to the shelf-monitoring program 120 along a second message path 135.
  • the shelf-monitoring program 120 then sends a return message to the planning module 125 along the second message path 135.
  • the planning module 120 next sends a message to an execution module 140 along a message path 145.
  • the planning module 125 and the execution module 140 can be components within a supply chain management application 150 or can be separate standalone components.
  • the execution module 140 After receiving the message from the planning module 125, the execution module 140 sends a first message to a manufacturing facility 155 along a message path 160 and a second message to a distributions center 165 along a message path 170. Based on this series of messages, the store is able to replenish its shelf with the item 105.
  • the LEWA 115 also can be used when the shelf is replenished.
  • the shelf-monitoring program 120 sends a message to the execution module 140 along the message path 135, which sends a message to the distribution center 165 along the message path 170. These messages are used to update the distribution center of the amount of stock on the shelf 105.
  • the distribution center can use the same IEWA or a separate IEWA to apply predictive models to the distribution center to estimate whether and when there will be an undesirable variation in inventory levels.
  • a store can use an IEWA to monitor the levels of an item on the shelf and in the inventory for one or more items to estimate potential undesirable variations in inventory levels.
  • the IEWA takes that sales data and uses algorithms to determine whether there will be an undesirable variation in the inventory levels and when to send an order to replenish the shelf and/or order more of that item from a warehouse or distribution center.
  • a warehouse or distribution center can use an LEWA to monitor the levels of an item within the warehouse, such as on shelves, on pallets, in quarantine, or at another location within the warehouse, to determine whether there will be an undesirable variation in the inventory levels and when to send an order to replenish.
  • Customers of the warehouse such as a retailer or a factory, order the item from the warehouse.
  • a consumer product goods (“CPG") retailer can order a pallet load of an item which the warehouse operator loads onto a delivery truck of either the warehouse, the retailer, or a third party logistics supplier.
  • the warehouse operator can use a wireless communications device to notify the inventory management software that a pallet-load of a particular item has been transferred from the warehouse.
  • CPG consumer product goods
  • Either the wireless communications device or the inventory management software can be programmed to notify the IEWA that a pallet-load of the particular item has been transferred from the warehouse.
  • the LEWA takes that transfer data and analyzes it using algorithms to determine whether there is likely to be an undesirable variation in inventory levels and when to order addition stock of that item.
  • One example of a framework of the LEWAs, an example of a messaging system used by the LEWAs, and the algorithms used by the IEWAs are described in more detail below.
  • an agent framework architecture 200 includes three agents 205, 210, 215, each of which communicates with an agent framework 217 that provides frameworks services: a message transport service ("MTS") 220, a directory facilitator ("DF") 225, an agent management service 230 ("AMS”), and a Post Office 235.
  • the agents 205, 210, 215 can be LEWAs.
  • the agent framework 217 refers to the container that provides programmatic support for agents, hi this context, agent framework 217 can be implemented in Java code that runs in a single virtual machine on a particular host. Additional services are used to support the agents running inside the framework. These additional services include life cycle management, directory services, communication, configuration, logging, persistence, and notification.
  • the MTS 220 is a service that is provided by a particular agent framework and allows agents to send messages, such as an AclMessage, to other agents.
  • An AclMessage encapsulates a communication between two agents and has some characteristics in common with an email message, such as, for example, specifying a sender and recipient, having a subject, and providing message content.
  • the DF 225 is a service provided by a particular agent framework and provides the framework's agents with access to the central directory service.
  • the directory service is a central service that provides identity (white page) and capability (yellow page) search facilities across an agent community. There is one directory service for one agent community.
  • a directory service might be federated with the directory service for another agent community.
  • An agent community is a collection of agent frameworks (and therefore agents) that collaborate to provide various business functions.
  • a community can consist of several agent frameworks, and each framework can in turn contain several agents.
  • the AMS 230 is a service provided by a particular agent framework and provides agent lifecycle management facilities within the framework.
  • the facilities allow remote management of a community of agent frameworks by essentially providing an external interface to the AMS in each agent framework.
  • the AMS allows administrators and control agents to manage the execution of other agents by stopping, starting and suspending agents.
  • the agent framework architecture 200 also includes an administrative user interface ("AUI") that allows a system administrator to manage an agent community.
  • the AUI uses the directory service to identify currently running agent frameworks. It then uses each framework's AMS to manage the agents within that framework.
  • the post office 235 is a service provided by a particular agent framework.
  • the post office 235 receives and maintains AclMessages addressed to the agents within that framework.
  • FIG. 3 illustrates the directory service 240 of an agent community 245 that includes three agent frameworks 217a, 217b, 217c, each of which can, for example, run on a separate host machine.
  • the directory service 240 is a centralized service that is used to unify the agent community 245.
  • the service can be implemented using the Java Naming and Directory Interface ("JDNI").
  • FIG. 3 shows only three agent frameworks 217a, 217b, 217c and, for simplicity, each agent framework contains only a single agent 205a, 205b, 205c, respectively. Nonetheless, the directory service 240 can be implemented with more agent frameworks and each agent framework can include more agents.
  • the directory facilitator 225 in each agent framework 217 updates the directory service 240, registering new agents as they are started up and deregistering agents as they are shut down.
  • the directory service 240 then allows agents to search for other agents, both by identity (for example, using a service that is analogous to a white page service) and capability (for example, using a service that is analogous to a yellow page service).
  • the DF 225 provides access to this functionality within each agent framework 217.
  • FIG. 4 illustrates the messaging architecture of the agent community 245.
  • the agent 205a requires the type of service provided by the agent 205b
  • the agent uses its DF function 225a to perform a capability directory lookup.
  • the DF 225a informs the agent 205 a that the agent 205b is capable of the type of service that it desires.
  • the agent 205 a then formulates an AclMessage and asks its local MTS 220a to deliver that message.
  • the MTS 220a uses the DF 225 a to locate the Post Office 235b for the agent 205b, and delivers the message to that Post Office 235b.
  • the messaging architecture can be implemented using Java's Remote Method Invocation ("RMI") to deliver the message to the Post Office 235b.
  • RMI Remote Method Invocation
  • the architecture can be implemented using other message transport mechanisms in the MTS, such as SOAP, JMS, CORBA, and JLNI.
  • FIG. 5 illustrates a communication sequence for a typical communication scenario 300 that is indicative of the type of interactions that occur within the agent community 245.
  • the agent 205a can be implemented as an inventory early warning agent that is used to monitor the level of inventory on a shelf or within a store or a warehouse, process the inventory information, determine whether to replenish the inventory, and communicate a request to replenish the inventory.
  • the agent 205a can receive inventory data from, for example, a computer system that receives sales data from a cash register scanning system.
  • the agent 205 a accumulates the sales data and processes the data continuously, periodically, or in a batch mode. As described in greater detail below, the agent 205 a also must analyze the data and determine whether to request a replenishment of the inventory.
  • the analysis can be as simple as a comparison of the amount of stock remaining to a threshold value or as complex as a statistical analysis of the likelihood that the stock will be emptied if the stock is not replenished within a set time.
  • the scenario of FIG. 5 assumes that the agent 205 a has determined that additional stock is required and will contact the agent 205b to deliver a replenishment request in a message.
  • the agent 205b can be, for example, an agent used at a manufacturer or warehouse to communicate inventory levels to agents at stores that are served by the manufacturer or warehouse.
  • the agent 205a wishes to send a message to agent 205b.
  • the diagram shows the sequence of messages between participating objects to accomplish this scenario.
  • the agent 205a invokes a searchDirectory method 305, which is processed by the DF 205 and the result is returned to the agent 205 a in a message 307.
  • the agent 205 a creates an AclMessage 310, retrieves the identification of agent 205b from the search result, sets the identification in the message's recipient field, and invokes a sendAclMessage 310.
  • the sendAclMessage 310 is processed by the message transport service 220a ("MTS"), which communicates with the DF 225a in a findPostOffice communication 315 to find the Post Office address 320 for the agent 205b.
  • the Post Office address 320 is communicated to the MTS 220a, which then delivers a message 325 to that Post Office 235b.
  • the agent 205b queries the Post Office 235b in a query 327 and then uses a retrieval 329 to retrieve the message from the Post Office 235b for processing.
  • the Post Office 235b also sends an acknowledgement communication 330 to the MTS 220a, which sends a communication 335 to the agent 205a.
  • the agent 205b can use an analytical method to determine whether to tell either the manufacturer to build more product or the warehouse to ship more product.
  • the agents discussed above can perform analysis of the inventory data that they receive to determine if there is likely to be a potential variation in inventory levels and, if so, when to request a replenishment of the inventory from a warehouse or the production of goods by a manufacturer to replenish the inventory in a warehouse or at a store.
  • the analysis can be a simple process involving the comparison of inventory on the shelf or in a warehouse to a threshold value. If the inventory level is above the threshold value, the agent does not request replenishment but if the inventory level is at or below the threshold value the agent requests replenishment.
  • Such an analysis does not take into account the likelihood that the stock on the shelf or in the warehouse will be sold quickly or slowly or that there can be some variability in the rate at which the stock is sold. Moreover, such an analysis does not take into account the likelihood that the warehouse or manufacturer will have sufficient stock on hand to supply the store, much less the likelihood that the warehouse or manufacturer can make a timely delivery.
  • the agent 205 uses a conditional probabilistic predictive statistical analysis to predict an expected inventory level, an upside confidence bound, and a downside confidence bound. For many companies, an expected inventory level that is near or below zero is undesirable. The company can determine the minimum low level, compare the expected inventory level to the company's minimum low level, and order more stock if these expected levels are below the minimum low level.
  • the input data for the conditional probabilistic predictive statistical analysis is forecast cumulative consumption, forecast cumulative replenishment, actual cumulative consumption, and actual cumulative replenishment.
  • FIG. 6 illustrates these values over time.
  • the input data for the analysis includes a time advance, an actual cumulative replenishment, a forecast cumulative replenishment, an actual cumulative consumption, and a forecast cumulative consumption.
  • the time advance variable is the setting describing how far in advance the prediction is intended. For example, a company concerned about its inventory levels for the next week would set the time advance to be seven days into the future.
  • the actual cumulative replenishment variable describes the total amount of new material actually received into inventory up until the time advance is reached.
  • the forecast cumulative replemshment describes the total amount of new material activities expected to be received into inventory up until the time advance is reached.
  • the actual cumulative consumption describes the total amount of material actually withdrawn from inventory up until the time advance is reached.
  • the forecast cumulative consumption describes the total amount of material expected to be withdrawn from inventory up until the time advance is reached.
  • the expected inventory level is the expected inventory level for the given time advance.
  • the upside 10% confidence bound is the inventory value that the inventory level has a 10% chance of exceeding.
  • the downside 10% confidence bound is the inventory value that the inventory level has a 10% chance of being below.
  • Any suitable statistical method can be used to determine the inventory level and confidence bounds.
  • the statistical analysis used by the agent can be implemented with a predictive statistical model that can predict the variability at any point along the planned inventory trajectory, can be started with very little historical inventory data, and can improve as the data is generated over time.
  • the plot of planned and actual inventory trajectories illustrated in FIG. 6 creates a number of records for training the predictive model: one for each time point in the plot. As such, such a dataset will only be complete when the data has been actually generated right up until the end of the plot.
  • One such statistical model that can be used is a probabilistic inference model based on a conditional Gaussian approximation. This model, its derivation, and its application to the inventory early warning agent are explained in more detail below.
  • the historical data is a record of planned and actual inventory levels over time. This data is generated as follows. Initially, at any given point in time, for example, day one, there is a forecast for consumption and replenishment for the next ⁇ time periods and thus a forecast for the actual inventory levels out to time ⁇ . Along with this predicted consumption there is assumed to be a planned replenishment resulting in planned inventory levels out ⁇ time periods into the future. The next day a new forecast is generated for consumption and replenishment out to time ⁇ + 1 resulting in another planned inventory. Using this historical data, D , the model is based on determining the density R(I
  • I a (t) be the actual inventory at time t > 0.
  • the inventory is not measured at all times, but can be interpolated from a piecewise constant function.
  • the breakpoints defining the actual inventory are given by (r(b),t"(b) for b e [l,R ⁇ ] where B" is the number of breakpoints defining the actual inventory and all t(b) > 0.
  • I" (•) and t" (•) $ are the B"
  • any time t there are a number of plans that predict the inventory at that time.
  • the number of plans forecasting any particular time is always
  • the model of inventory levels is based on scenarios to which the inventory is actually subjected. As such, there are a variety of occurrences that might happen to the inventory levels between measurements.
  • the simplest case is that the inventory was unchanging between measurements. This is termed the piecewise constant case since it is possible to interpolate inventory levels between measurements by using a piecewise constant function.
  • I(t) I 1 -t-(I 2 -I 1 )(t-t 1 )/(t 2 -t 1 ).
  • the model is based on the assumption of a piecewise constant rather than piecewise linear interpolation, although either choice or other interpolation mechanism is acceptable for data preparation since the algorithm is independent as to the choice. All the algorithm requires is that the inventory levels be measured at equally spaced intervals, ⁇ , and this can always be accomplished by interpolation by either method.
  • the particular approach used with the IEWA uses the essential idea of exploiting correlations in errors between planned inventories and the actual inventories themselves. These correlations are used to estimate a probability density for a number of errors and then condition this density on the available data at the time a prediction needs to be made to find the likely errors at the time of prediction. The most likely correction to the planned inventory can then be applied to estimate the most likely actual inventory. The full probability can be used to define confidence intervals around this most likely prediction. Examples of correlations in errors include seasonal errors in inventory, such as season changes in inventory related to holidays or other seasonal events.
  • the next step is to estimate R(f health) given the historical data by exploiting two types of correlations that likely exist in the data.
  • the elements of f detox are likely correlated because if inventory is high (or low) at time nA + ⁇ x then it is also likely to be high (or low) at the later time nA + ⁇ 2 so long as ⁇ 2 is not much greater than ⁇ ! .
  • the plans formed at subsequent planning periods will be correlated since the more recent plan is probably an update of an older plan.
  • the algorithm uses a model of the joint density P f n ,f JI _ 1 ,-",f n _ p ) .
  • a suitable model for the joint density is that of a Gaussian joint density, as follows:
  • this Gaussian model can be used for predicting inventory levels and calculating the mean and covariance.
  • the first step is integrating to determine P(f n
  • PQ I fate ⁇ - - ) J4C ⁇ ⁇ ⁇ ⁇ df _ p P ⁇ i n _9" ⁇ t p n _ p ) (0.2)
  • P ( ⁇ n , f n ⁇ _ j , f _ ⁇ ⁇ ⁇ ⁇ , f _ p , f ⁇ ) is given by Eq. (0.1 ) .
  • the integration results in another Gaussian distribution with mean ⁇ ⁇ and covariance ⁇ ⁇ .
  • the notation a(N r ) indicates the
  • A(N r ,N c ) is the
  • gives the number of elements in the set or vector S .
  • the complete algorithm can be used to make the prediction as follows: (1) assuming a value of p (or having the algorithm determine p ), construct an estimate of ⁇ and ⁇ from the historical data; (2) construct the index set N ⁇ and form ⁇ ⁇ and ⁇ ⁇ from ⁇ and ⁇ according to Eq. (0.3); (3) using historical error values construct the vectors of past errors ⁇ _ x ,- --,i ⁇ _ p ; (4) build ⁇ (fgate ⁇ ) and ⁇ (fgate ⁇ ) according to Eqs.
  • the first two steps can be done offline in a discrete calculation since they have no dependence on the time for which the prediction is being made.
  • To accomplish the first step and determine the parameters of the Gaussian there a number of approaches or models that can be used.
  • a first model is based on frequency counts while a second and a third method use an autoregressive model of the multivariate time series ⁇ f ⁇ .
  • the second and third models differ in the underlying simplifications made, with the second model being more accurate but more complex.
  • the model is adaptive to and learns from changes in conditions.
  • the model uses a simplification in order to tractably model P( ⁇ ford_ v - • -,icot_ p ) , which
  • the remaining probability P(f ) is
  • ⁇ ' can be inverted using the identity that if
  • each conditional probability can be modeled as an autoregressive process of successively smaller orders.
  • the parameters of each process can be determined from the historical times series.
  • the model uses the same Gaussian assumption as before.
  • Each conditional probability for the AR(p -i) process is written as
  • Gaussian with precision (inverse covariance) given by ⁇ ⁇ ! (whose rows and columns have been ordered according to f ), which can be formed as
  • This matrix can be easily implemented in software as executable instructions to provide the covariance.
  • the mean is determined by completing the square:
  • the algorithm is computationally very efficient.
  • the covariance matrix and its mean can be estimated with a single pass through the data by recording frequency counts. If there zreD data points then the complexity of accomplishing both tasks is of the order 0(D(p + V)q((p + ⁇ )q +1)) . Prediction is not much more complicated. The dominant factor is the computation of ⁇ "1 , which is of order 0( ⁇ N 2 ⁇ 3 ) . Moreover, this matrix inverse can be stored and need not be recalculated with each prediction.
  • the above algorithm must be modified so that it can operate in an online mode in which all parameters of the algorithm (that is, the mean and covariance) are continually updated as new data arrives so that it is adaptive.
  • the algorithm also can be modified to discount data from the past by a factor 0 ⁇ ⁇ ⁇ 1. This allows the algorithm to learn from recent events and adapt to changing conditions.
  • ⁇ ⁇ - ⁇ + ⁇ D ⁇ - l ( - ⁇ ⁇ ⁇ x (0.13) ⁇ D ⁇ ( ⁇ ) D ⁇ ( ⁇ ) and
  • Another way to set ⁇ is by minimizing the cross-validation error.
  • the IEWA uses a probabilistic predictive algorithm embodied in executable instructions on one or more computer systems to estimate future inventory levels as follows: First, the IEWA takes any historical inventory data to calculate a Gaussian distribution with a mean and covariance and provide an inventory prediction at a set time in the future (step 405). The computer systems can be further programmed to compare the prediction to a first company-specified low inventor threshold and to a second company-specified high inventory threshold (step 410). The LEWA can be programmed to order additional stock for inventory if the predicted level is below the low inventory threshold (step 415) or cancel existing replenishment orders if the threshold is above the high inventory threshold (step 420).
  • the IEWA then records time dependent data related to inventory consumption and replenishment (step 425) so that it can calculate time dependent cumulative forecast consumption, cumulative forecast replenishment, cumulative actual consumption, and cumulative actual replenishment (step 430).
  • This data can vary over time, seasonally, and for unknown reasons, such as recessions or other changes in economic conditions.
  • the JJEWA uses this new data with the historical data to update the Gaussian distribution, mean, and covariance and calculate a new predicted inventory level for the next time period (step 435). Over time, the data accumulated has the likelihood of improving the predicted inventory level within the supply chain management system as the amount of data increases in volume.
  • the IEWA can use two modes of operation for the predictive function: the aggregate mode and the detailed mode.
  • the aggregate described above, generally uses real time data and accumulated data.
  • the detailed mode is based on a detailed model that used detailed data that has been previously generated.
  • the prediction task is split into two agents: (1) an Activity Risk Agent, and (2) an Early Warning Agent.
  • the Activity Risk Agent conducts activities learning about potential variation based on historical performance, determines estimates of variation for currently planned activities, and provides those estimates to the Early Warning Agent.
  • the Activity Risk Agent uses individual activities records for the activities learning about potential variation.
  • the activities records contain both promised and actual delivery time and quantity. Using this data, the detailed model generates predictions for actual delivery dates and quantities for not yet completed activities.
  • the detailed model uses decision trees as the initial approach to machine learning.
  • the detailed model uses Classification and Regression Trees (CART®) because these trees: (1) are easy to train and maintain, and (2) perform nearly as well as other machine learning techniques, such as neural networks, that are usually much harder to train, and from which knowledge cannot be extracted easily.
  • classification and regression trees can be implemented and applied to data that consists of both continuous and discrete attributes. However, the use of other machine learning techniques is not precluded by this choice.
  • the general assumption upon which the detailed model is based is the observation that different suppliers or consumers of a particular material can have behavioral patterns that determine how much and when they deliver or consume materials with respect to what they promise in their order. For example, some suppliers can follow their delivery schedule very closely, some suppliers can be consistently late in delivering their orders, and some consumers can cancel orders frequently. These behavioral patterns can be learned if they are statistically significant. For example, some consumers, such as retailers, can regularly cancel orders after holidays because of prior overstocking of their inventories in anticipation of unrealized sales. Similarly, some suppliers can regularly deliver orders late around holidays because, for example, they regularly over-commit to their customers. Decision trees can be used to leam these patterns.
  • an algorithm used in the detailed model involves the building of a decision tree, determining splitting points on the decision tree, pruning the decision tree, extracting precise measurements of variability from the decision tree, and date pnming.
  • a decision tree predicts an attribute by asking a succession of yes/no questions about known predictors. Each non-leaf node in the tree contains a question, and each leaf node contains a prediction.
  • Such predictions can be, for example, single values plus standard deviations when the response is a continuous variable, or a class annotated with a probability when the response is a discrete valuable (that is, known as a classification tree).
  • the questions in non-leaf nodes refer to values of predictor variables. When those variables are continuous, the questions are in the form “is x less than C?" where C is a constant value. When the variables are discrete, the questions are in the form "does x belong to one of the classes A, C or F?" where A, C and F are some of the classes to which the discrete variable can belong.
  • the task of extracting a prediction from a decision tree can be relatively simple, as it involves traversing the tree according to the questions the tree poses. Building the tree itself, however, can be quite difficult.
  • data-mining techniques are used that infer a tree from detailed data where the response variable is known.
  • these techniques seek to find splitting points in the predictors (that is, these techniques seek questions) such that the split is maximally informative, thus minimizing the number of subsequent questions needed to reach a prediction.
  • decision tree is the CART®, noted above.
  • This type of decision tree advantageously can be rapidly inferred from data that can contain both continuous and discrete predictors. Moreover, the response variable can be continuous or discrete itself.
  • Classification and regression trees are constructed in two stages: (1) a growing stage and (2) a pruning stage. In the first, growing stage, the data is recursively split using a criterion that makes the split maximally informative. In the second, pruning stage, the tree is pruned back to avoid over-fitting and to obtain an adequate level of generality.
  • the growing part of the algorithm works according to the following general procedure. Initially, for all predictors, select those predictors in which the data exhibits the possibility of an informative split. If, however, no predictors meet this criterion, exit. Otherwise, choose the split that is most informative, and repeat until no more splits are considered worthy.
  • the criterion used for seeking split points when the response variable is continuous is based on Analysis of Variance (ANONA) techniques. Using A ONA, if the predictor variable x itself is continuous, it is sorted in ascending order, and then a point x s is chosen such that the expression
  • the Analysis of Variance approach also can be used for discrete predictor variables.
  • each of the classes that the predictor variable can belong to is ranked by the mean of their corresponding response values, and then a split is chosen using the same criterion as above by shuffling classes between two groups starting with those of smaller rank and moving them in a Graycode order.
  • the Gini rule When the response is discrete, the Gini rule is used, which requires the searching for a splitting point that isolates most frequently the largest class in the response. For this purpose, assume that the predictor is continuous, and that, as above, the splitting point is x s . This point partitions the response into two sets, P L and P R . For each response class j, define p(j ⁇ IX) and p(j ⁇ R) , the probabilities that a member of classj belongs to P L or P R , respectively, as
  • the expression to minimize is the total impurity, given by PU I ⁇ C; I * ), where I(p) is the impurity function, and can be defined either as 1-p 2 or - p g(p).
  • the procedure is similar when the predictor is discrete, in which case the split point is selected using Graycode-ranking.
  • the optimal size of a decision tree is inferred by cross-validation.
  • the training data is split into multiple equal-sized buckets, and separate trees are grown for subsets of the data that exclude one of the buckets respectively.
  • Each tree is then pruned at all possible degrees, and for each degree, the tree's performance is evaluated using the bucket that was excluded in its construction.
  • the best degree of pruning is then chosen as the one that maximizes average performance across the data in all held-out-buckets.
  • the original tree is then pruned to the chosen degree.
  • predictions extracted from decision trees are parsimonious, typically consisting of a single-valued prediction and a confidence interval.
  • more detailed information can be extracted by annotating each leaf of the tree with the values of the predictor variable of the training data that correspond to that leaf of the tree.
  • the tree then can be seen as a partition of the set of training values for the predictor, and a prediction consists of an entire element of this partition. If binned, such element constitutes a histogram. By performing statistics on the histogram, a user can restore the original parsimonious predictions.
  • the user obtains the following data or fields to create decision trees. From this data, three models are extracted: (1) a model to predict variability on actual delivery quantity; (2) a model to predict variability on actual delivery date; and (3) a model to predict order cancellation probability.
  • the dates should be encoded into multiple variables as day of the week, day of the month and month of the year, in addition to a single numeric encoding that represents absolute time.
  • the data or fields and an explanation of that data or field are as follows:
  • Each of the above fields can be represented as one or more numerical or categorical features for use in the machine-learning algorithm (for example, the expected delivery date can be expressed as day of week, day of year, month of year, etc.).
  • a Data Collection Agent must construct these records by matching up planned activities with their outcomes.
  • the "source” for an activity is the partner or resource involved in that activity.
  • the source is as follows: (1) for replenishment of raw materials, the source is the supplier; (2) for consumption of raw materials, the source is a resource (for example, production line); (3) for replenishment of finished products, the source is a resource (for example, production line); and (4) for consumption of finished products, the source is a customer. If necessary, additional fields can be implemented for use in the special circumstance of predicting the arrival of last minute orders to be able to predict the variation resulting from orders that are not yet known but that might arrive.
  • the absolute-time encoding of dates does not play a significant role in prediction since all planned orders used for prediction will lie in the future with respect to historical orders used to build the tree. Instead, if the tree contains splits on absolute time, this will be used as an indication that a qualitative shift in behavior has occurred at those points in time. Training data that is older than the most recent of such points is discarded and a new tree is built using only the most recent data. However, if splits on absolute dates occur deep in the tree, then data should only be discarded if the data reaches those deep splits.
  • the detailed model described above allows the user to annotate planned activities with potential variation, which, for example, allows the user to annotate a sequence of planned activities with potential variation.
  • a company in a supply chain can obtain the probability distributions of various inventory levels across time. Ln general, at an initial time an inventory contains a known number of items. In the subsequent time interval, there is scheduled a number of replenishment operations which are planned to be a certain time and quantity, but there is a stochastic variation expected in both time and quantity for each operation. It is assumed that both time and quantity are made discrete so that the distributions are discrete and specifiable as a histogram.
  • the variation in time and quantity is assumed to be independent so that they can be completely specified by two independent histograms over one variable only.
  • one consideration to be analyzed is how to efficiently propagate the probability distribution between the end of time interval t and the end of time interval t + 1. To analyze this efficiency consideration, there are a number of observations that are helpful.
  • this operation has computation order 0(lm) where / and m are the numbers of nonzero elements in f t+ ⁇ and P t .
  • the detailed predictive model can be further developed using time evolution filters for various operations. For example, "ordinary” (that is non-stochastic) transactions can be put into the equations described above. First, if no transaction occurs in the interval t, the time evolution filter (TEF) is given by
  • this TEF will not have connected support, but will be non-zero both at 0 and in a region around the expected transaction quantity. This will still allow for sparcity considerations but more care must be taken to track the support of the TEF.
  • the TEF divides into four (possibly overlapping) regions: one region for when neither occurs, two regions for when one occurs and not the other, and one region for when both occur.
  • the possibility of multiple transactions occurring in the same time interval complicates the spread of the probability distribution over quantity.
  • TEF is equal to the complete TEF for the transaction (as if there was no time variation) as follows:
  • T is the set of all transactions and K is the convolution series operator.
  • the computational complexity of this algorithm depends on several factors, including: (1) n, the number of bins the probability histogram is to be evaluated at for each time; (2) T, the total time duration of the probability distributions for all transactions; (3) m, the number of transactions; and (4) ⁇ Q>, the average support length of the quantity distribution for the transactions. Since each transaction is active for a number of time steps equal to the duration of its time distribution, a total of T convolutions with Pt are completed for all transactions. Further, since each transaction is incorporated exactly once into P0, an additional m convolutions are performed.
  • Each convolution is completed between a vector of average length ⁇ Q> and a vector of length n (P0 or Pt) which takes on the order of ⁇ Q>n operations.
  • This order of operations is based on assuming: (1) that the single element indicating probability of not happening is inconsequential and can be ignored; (2) that no advantage is taken of possible sparseness of P0 or Pt; and (3) there is no use of a fast Fourier transform technique. If the user assumes that the convolutions are the most computationally intensive operations, the result is a computational order of (T+ m) ⁇ Q>n.
  • TEFs have connected support
  • the support of the distribution will continue to be connected for all time and this will be as efficient as possible.
  • the support of the distribution will become fragmented and there will still be zeros between the upper and lower bounds.
  • An example of TEFs failing to have connected support includes the situation in which there is a non-zero probability of failure.
  • a more thorough elucidation of the support of the distribution that expresses the support as a union of intervals would ameliorate this problem, but the computational effort of keeping track of a potentially exponentially increasing number of intervals and deciding when two intervals overlap to merge them would likely use as much computing resource as it saves.
  • a second method of speeding up the computation is to have a convolution that checks whether or not an element of the probability distribution is zero before factoring in the convolution. This type of check likely would save many operations in which numbers are multiplied by zero.
  • a third method of speeding up the computation takes advantage of the associative nature of the convolution operator. Ln the event that more than one transaction occurs in a time interval, the individual TEFs for each transaction can be applied individually. Because the support of the TEF for a single transaction has at most two components, a sparse representation can be easily used.
  • a fifth method of speeding up the computation uses an alternative approximation method that eliminates small probabilities at each iteration, either by setting a threshold on the smallest probability to be considered or by only keeping the highest N elements for some N. Again, the increase in computation speed results from the faster convolution that is a consequence of having fewer non-zero elements in the quantity distribution.
  • Any of these five methods of speeding up the computation can be implemented to conserve computation resources. Moreover, one or more of these methods can be combined to further speed up the computation.
  • bicumulative display In addition to the computation of the probability distribution of inventory over time, it is also necessary to consider appropriate graphical methods to display these results to an end user. There are at least two methods of display: (1) bicumulative display and (2) relative density display.
  • the bicumulative display method takes the computed probability density function (pdf), sums to form the cumulative distribution function and applies the function f ⁇ ⁇ x-» 2xH x + 2(l-x)H x — 2 , V 2y where His the ⁇ eaviside function, to the cumulative distribution function. The result is then plotted as an image intensity with 1 being the darkest and 0 being the lightest.
  • This method has the advantage of highlighting the median value of the distribution, but can disadvantageously hide any multi-modality in the distribution that can result from orders having significant probabilities of cancellation and give a false impression of a single- humped distribution.
  • the relative density display method uses the PDF only and normalizes at each time step so that the inventory with the largest probability of occurring is assigned a renormalized probability of 1.
  • the function is then displayed as in the bicumulative display.
  • the renormalization ensures that at each time step, there is at least one inventory level shown at the darkest color so that the distribution does not "wash out” as the distribution spreads over time.
  • the relative density display method is more intuitive in that a dark segment of the plot indicates an increased probability of being at that location, and multi-modality is shown as more than one dark region at a single time step.
  • the normalization makes it more difficult to judge differences in probability between time steps, while differences in probability at different inventories at the same time step are easier to judge.
  • Another aspect of the detailed model that must be considered is the calibration of the confidence bounds that relate to inventory levels.
  • Supply-chain control systems tend to be self-correcting. For example, if inventory levels are too low, new sources of supply will be sought, or some demand will be cancelled or postponed. In fact, the Inventory Early Warning Agent is intended to enhance this process of self-correction. For this reason, it is expected that the confidence bounds of a model that inherently assumes a non-correcting system will be too wide. As such, predicted confidence bounds are corrected through scaling to compensate for the self-correcting nature of the supply-chain control system. To make this correction, histograms of actual variation are compared with histograms of predicted variation, a scaling factor that maximizes their similarity is derived, and that scaling factor is used for future predictions.
  • the number of records of replenishment or consumption activities can be extremely large (for example, tens of thousands) or extremely small (for example, two or three). Large quantities of data can require excessive processing time, and small quantities of data can provide insufficient records for learning. Both of these extremes, however, can be adequately handled by applying different techniques based on the size of the data set. For example, if the historical data set is too large, the model can use sampling from the data set. The unused part of the large data set can be later used for validation purposes. If the historical data set for a particular material is too small, it can be extended with records for other materials that involve the same partner or resource.
  • the model can be programmed to, for example, have a threshold that determines whether the data set is too large or too small and automatically samples the data or extends the records, respectively.
  • the model must also be able to handle activities that have multiple updates. For example, one particular activity can be updated many times before being completed, as in the case of a completion time for providing a good.
  • the completion time can be set initially, and then changed multiple times based on changes in, for example, production schedules and transportation logistics. Each time an activity is updated, a new record for machine learning is generated. If records are considered as independent entries for the model-building algorithms, then activities that are updated multiple times will weigh more heavily than activities that are updated rarely and cause inaccuracies in the model. To compensate for the over-weighing of certain activities, records are weighted by the fraction of time they persist before being updated or fulfilled. The fraction is computed relative to the total time elapsed between when the activity was first planned and when it was completed. Thus all activities will have the same weight, independent of the number of updates.
  • Predictor variables can be expressed as absolute deviations, percentages, or absolute values.
  • expected variation in quantity can be expressed in terms of units of the variable or in terms of percentage of the order.
  • the learning algorithm of the detailed model can be implemented to construct models that use different forms of the predictor and automatically select the one that performs better at prediction.
  • Decision trees can only predict a single variable at the time.
  • the detailed model involves the prediction of two or three variables. Since the model will be used to predict actual delivered quantities, actual delivery dates, and completion time spread, one approach is to build separate models, one to predict each of these variables. This approach, however, has the implicit assumption that these variables are independent. Nonetheless, this is one approach that can be implemented in the detailed model.
  • a second approach is to develop a model for one of variables to be predicted, and then use it as a feature in a model for one of the other variables to be predicted. Again, this approach can be implemented in the detailed model. These two approaches imply that there are at least two possible ways to assume dependence.
  • Some of the activities that relate to inventory management can involve splitting or combining activities. For example, one item, represented as a first activity, can be incorporated into a second item, and the resulting combined item delivered to inventory. Such an activity involves combined activities.
  • a splitting activity a product line can make items that are delivered to, for example, the work-in-progress inventory of multiple factories operated by the company. Such an activity involves splitting activities.
  • split activities can be treated as separate activities whose characteristics are prorated.
  • Combined activities can be treated either as single activities that are aggregated, or as independent activities that happened to be delivered at the same time.
  • Each of these methods can be more appropriate in some circumstances than other methods.
  • the Data Collection Agent described above with respect to obtaining data for constructing the decision frees, will additionally have the task of extracting clean records from the stream of messages pertaining to the splitting and combining activities.
  • the Activity Risk Agent must be able to model this type of situation to accurately represent typical supply chains. If there is a record of the four plan dates pertaining to activities (for example, creation date, current date, planned execution date, and actual execution date) then there is sufficient data to work out the probabilities of activities being created at the last minute. Using this information, the Activity Risk Agent can create new activities with the appropriate cancellation probability to represent last-minute orders (for example, the probability of the order being cancelled is equal to one minus the probability of it being created).
  • the Activity Risk Agent builds a decision tree that estimates the potential variation in outcome for a particular planned activity.
  • the decision tree can be converted into a set of human-readable rules that can be interpreted and possibly changed by the user.
  • the model-building algorithm used to create the Activity Risk Agent can be "seeded" with known rules about the variation of activities (for example, the user can modify the algorithm to include the condition that shipments of a particular material from a certain supplier arrive within 2 hours of the planned time).
  • Advantages of adding this ability for the user to inspect, change, and seed the learning algorithms are that it gives the user a greater understanding of the system and the ability to enter knowledge that can not be otherwise available.
  • the Activity Risk Agent optimally is based on knowledge of the relationship between the real-time end-product demand signal and demand from immediate customers at the location. To obtain this knowledge, a Real-Time Demand Forecast Agent takes the point-of-sales data information immediately when it is generated and evaluates how that will affect future orders. The Real-Time Demand Forecast Agent optionally can evaluate already planned orders to determine whether there are possible changes to the orders. Thus, the Activity Risk Agent uses data obtained from the Real-Time Demand Forecast Agent to forecast orders (or evaluate possibility of changes to planned/forecasted orders) at some future time t ' , given the current point-of-sales and other data, such as inventory levels, orders already in transition, etc.
  • a first scenario is that of a retailer-manufacturer relationship in which orders are placed by a store chain corporate center to a manufacturer.
  • a second scenario is that of a distribution center-manufacturer relationship in which orders are placed by a distribution center to a manufacturer as a result of aggregated/accumulated demand from individual stores.
  • a third scenario is that of a store-distribution center relationship in which orders are placed by an individual store to a distribution center. The third scenario is only relevant if there are time delays in order placement or deliveries between stores and distribution centers.
  • the information available to the decision maker can differ among the scenarios described above. As such, different algorithms can be prepared that are applicable to all of the above scenarios and other algorithms can be prepared that are better suited to one or two of the scenarios. Examples of these algorithms are described below.
  • the algorithms each include some common terms, as defined below: • time t e ⁇ 1,...,T ⁇ ; • i t - inventory at the beginning of period (day) t (measured after this day's shipment arrives);
  • the algorithms formulated below are based on the assumption that future orders are driven by: (1) inventory; (2) sales at time t and at prior times; and (3) by the orders already in the supply chain.
  • the algorithms also assume that there is a built-in time-delay in the existing supply chains (as illustrated in FIGs 8A and 8B). Because of this time delay, the information set available to a decision maker when placing an order at time t (for example on behalf of a group of stores), consists of the following:
  • I° ⁇ x r , ⁇ ⁇ t - ⁇ x ⁇ , ⁇ i ⁇ , ⁇ ⁇ t - ⁇ i ⁇ t ⁇ s r , ⁇ ⁇ t - _d. ⁇ , ⁇ o ⁇ , ⁇ ⁇ t - ⁇ t 0 ⁇ , where ⁇ t y (assumed to be a positive integer) is the information delay for the variable y .
  • the information set for the Real-Time Demand Forecast Agent is expressed as following:
  • the agent's goal is to "extract" the decision maker's ordering policies from the available data and then to "make” ordering decisions before they are actually made by the decision maker.
  • This ordering policy "extraction” can not always be optimal, as illustrated by the following scenario. Assume that the user, agent, or decision maker attempts at time t to "predict" what order will be placed by a decision maker at time t+ ⁇ t. Depending on the exact value of the ⁇ t, the information set available for the estimation of future orders, Y t , can not contain all the information in _h.el r +&i . Therefore, even if the ordering rule was extracted perfectly, the estimate can differ from the actual decision made.
  • one method that can be used is to estimate a set of functions f t+ t (I t r ), At e ⁇ 1,...,T ⁇ , each of which minimizes an appropriately selected distance function d between our approximation and the actual order: d(f M (I r ) — o M ) .
  • the distance minimization criterion of the distance function has been arbitrarily chosen here for simplicity but can be enhanced as necessary.
  • an order placed at time t can be a composition of orders resulting from both of the above processes.
  • stores receive information about forthcoming promotions (or perhaps make promotional decisions) and can order ahead of time.
  • forthcoming promotional information is unlikely to be provided by the agent that obtains the actual data. If this assumption is not correct and information about forthcoming promotions is received, then constructing the model or simulation becomes simpler.
  • One particular method that stores use to make decisions about the quantity to be ordered at the beginning of the day t is keep the expected inventory at the beginning of the day t+l (after I's day shipment arrives) at a value or level that is equal to the value of the model stock m :
  • s t m + f. t+1 _ x -o t -i t , (0.17) where f t:t+! _ x is the expected total consumption in the periods from t to t + l-1.
  • an amount of goods ordered at a particular time t + 1 is a linear combination of an earlier order and sales.
  • an order at time t+1 can be expressed as a linear combination of point-of-sale, order and inventory levels from previous periods. This provides a basis for using the linear Real-Time- Demand prediction algorithm, described below.
  • Another possible method to correlate the supply chain data described above is to assume that the time-series of orders and of other variables is auto-correlated.
  • Such an auto-correlation is illustrated by the orders autocorrelation function for a simulated data set in FIG. 8C.
  • Matlab® and S-Plus® algorithms can be used that will fit a given time-series to an auto-correlated process. As such, the fit can be provided by an auto-correlation or a linear regression.
  • linear regression algorithms instead can be used in the model in response to difficulties with the algorithms described above.
  • one difficulty in implementing the linear regression algorithms is the necessity of taking into account seasonal effects in the consumer demand. For example, the day of the week effect is often pronounced in the real-world supply chains.
  • a classification and regressions decision tree (CART®) algorithm can be used to take such seasonality into account. Otherwise, the set of predictor variables is similar to that described for the linear algorithm above.
  • Another algorithm that can be used is the cross-correlation method.
  • the algorithms that are used can be selected such that they do not take inventory into account at all.
  • One algorithm that is appropriate in such a situation is the cross- correlation method of the instrument function estimation.
  • the invention can be implemented in digital electronic circuitry, or in computer hardware, firmware, software, or in combinations of them.
  • the invention can be implemented as a computer program product, i.e., a computer program tangibly embodied in an information carrier, e.g., in a machine-readable storage device or in a propagated signal, for execution by, or to control the operation of, data processing apparatus, e.g., a programmable processor, a computer, or multiple computers.
  • a computer program can be written in any form of programming language, including compiled or interpreted languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.
  • a computer program can be deployed to be executed on one computer or on multiple computers at one site or distributed across multiple sites and interconnected by a communication network.
  • Method steps of the invention can be performed by one or more programmable processors executing a computer program to perform functions of the invention by operating on input data and generating output. Method steps can also be performed by, and apparatus of the invention can be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).
  • FPGA field programmable gate array
  • ASIC application-specific integrated circuit
  • processors suitable for the execution of a computer program include, by way of example, both general and special purpose microprocessors, and any one or more processors of any kind of digital computer.
  • a processor will receive instructions and data from a read-only memory or a random access memory or both.
  • the essential elements of a computer are a processor for executing instructions and one or more memory devices for storing instructions and data.
  • a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks.
  • Information carriers suitable for embodying computer program instructions and data include all forms of non- volatile memory, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DND-ROM disks.
  • semiconductor memory devices e.g., EPROM, EEPROM, and flash memory devices
  • magnetic disks e.g., internal hard disks or removable disks
  • magneto-optical disks e.g., CD-ROM and DND-ROM disks.
  • the processor and the memory can be supplemented by, or incorporated in special purpose logic circuitry.
  • the invention can be implemented in a computing system that includes a back-end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front-end component, e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the invention, or any combination of such back-end, middleware, or front-end components.
  • the components of the system can be interconnected by any form or medium of digital data, communication, e.g., a communication network. Examples of communication networks include a local area network ("LAN”) and a wide area network (“WAN”), e.g., the Internet.
  • LAN local area network
  • WAN wide area network
  • the computing system can include clients and servers.
  • a client and server are generally remote from each other and typically interact through a communication network.
  • the relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Marketing (AREA)
  • Development Economics (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Educational Administration (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • General Factory Administration (AREA)

Abstract

Cette invention concerne des procédés et un dispositif, dont des produits de programmes informatiques, et des techniques de mise en oeuvre et d'exploitation pour la surveillance et la prévision de diverses données relatives à un stock. Un dispositif de traitement de données interagit avec un ou plusieurs ordinateurs commerciaux pour l'obtention de données sur des activités soit de réapprovisionnement, soit de consommation relatives à un stock. La mise en place d'un modèle pour variations de stock permet d'établir des estimations relatives à des variations soit d'activités actuellement prévues de réapprovisionnement des stocks, soit d'activités actuellement prévues de consommation. On utilise ces estimations pour prévoir des futurs niveaux de stock. Sont également décrits un ou plusieurs ordinateurs d'agent.
PCT/US2002/036813 2001-11-14 2002-11-14 Agent utilisant un modele predictif detaille WO2003042793A2 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
AU2002346417A AU2002346417A1 (en) 2001-11-14 2002-11-14 Agent using detailed predictive model

Applications Claiming Priority (8)

Application Number Priority Date Filing Date Title
US33622701P 2001-11-14 2001-11-14
US60/336,227 2001-11-14
US38463802P 2002-05-31 2002-05-31
US60/384,638 2002-05-31
US10/208,186 2002-07-31
US10/208,186 US7222786B2 (en) 2001-11-14 2002-07-31 Inventory early warning agent with correction by error correlation calculation
US10/281,074 US7644863B2 (en) 2001-11-14 2002-10-24 Agent using detailed predictive model
US10/281,074 2002-10-24

Publications (2)

Publication Number Publication Date
WO2003042793A2 true WO2003042793A2 (fr) 2003-05-22
WO2003042793A3 WO2003042793A3 (fr) 2004-07-15

Family

ID=27498717

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2002/036813 WO2003042793A2 (fr) 2001-11-14 2002-11-14 Agent utilisant un modele predictif detaille

Country Status (2)

Country Link
AU (1) AU2002346417A1 (fr)
WO (1) WO2003042793A2 (fr)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2005052825A2 (fr) * 2003-11-20 2005-06-09 Roland Mary Systeme de gestion et de controle automatique de l'utilisation de ressources humaines et materielles
CN111160706A (zh) * 2019-07-16 2020-05-15 华为技术有限公司 模拟供应链风险响应的方法、装置和系统
CN112270586A (zh) * 2020-11-12 2021-01-26 广东烟草广州市有限公司 一种基于线性回归的遍历方法、系统、设备和存储介质
US11042837B2 (en) 2018-09-14 2021-06-22 Walmart Apollo, Llc System and method for predicting average inventory with new items
CN114048931A (zh) * 2022-01-13 2022-02-15 北京京东振世信息技术有限公司 补货信息生成方法、装置、电子设备和计算机可读介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5615109A (en) * 1995-05-24 1997-03-25 Eder; Jeff Method of and system for generating feasible, profit maximizing requisition sets
US5946662A (en) * 1996-03-29 1999-08-31 International Business Machines Corporation Method for providing inventory optimization
US6205431B1 (en) * 1998-10-29 2001-03-20 Smart Software, Inc. System and method for forecasting intermittent demand
US6609101B1 (en) * 1999-03-26 2003-08-19 The Retail Pipeline Integration Group, Inc. Method and system for determining time-phased product sales forecasts and projected replenishment shipments for a retail stores supply chain
US6622056B1 (en) * 2000-11-17 2003-09-16 Johan Lindell Method and system for supply chain or supply network control

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5615109A (en) * 1995-05-24 1997-03-25 Eder; Jeff Method of and system for generating feasible, profit maximizing requisition sets
US5946662A (en) * 1996-03-29 1999-08-31 International Business Machines Corporation Method for providing inventory optimization
US6205431B1 (en) * 1998-10-29 2001-03-20 Smart Software, Inc. System and method for forecasting intermittent demand
US6609101B1 (en) * 1999-03-26 2003-08-19 The Retail Pipeline Integration Group, Inc. Method and system for determining time-phased product sales forecasts and projected replenishment shipments for a retail stores supply chain
US6622056B1 (en) * 2000-11-17 2003-09-16 Johan Lindell Method and system for supply chain or supply network control

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2005052825A2 (fr) * 2003-11-20 2005-06-09 Roland Mary Systeme de gestion et de controle automatique de l'utilisation de ressources humaines et materielles
WO2005052825A3 (fr) * 2003-11-20 2005-07-21 Roland Mary Systeme de gestion et de controle automatique de l'utilisation de ressources humaines et materielles
US11042837B2 (en) 2018-09-14 2021-06-22 Walmart Apollo, Llc System and method for predicting average inventory with new items
CN111160706A (zh) * 2019-07-16 2020-05-15 华为技术有限公司 模拟供应链风险响应的方法、装置和系统
CN112270586A (zh) * 2020-11-12 2021-01-26 广东烟草广州市有限公司 一种基于线性回归的遍历方法、系统、设备和存储介质
CN112270586B (zh) * 2020-11-12 2024-01-02 广东烟草广州市有限公司 一种基于线性回归的遍历方法、系统、设备和存储介质
CN114048931A (zh) * 2022-01-13 2022-02-15 北京京东振世信息技术有限公司 补货信息生成方法、装置、电子设备和计算机可读介质

Also Published As

Publication number Publication date
WO2003042793A3 (fr) 2004-07-15
AU2002346417A1 (en) 2003-05-26

Similar Documents

Publication Publication Date Title
US7644863B2 (en) Agent using detailed predictive model
US7222786B2 (en) Inventory early warning agent with correction by error correlation calculation
US10997614B2 (en) Flexible feature regularization for demand model generation
US7398232B2 (en) Inventory early warning agent in a supply chain management system
US10936947B1 (en) Recurrent neural network-based artificial intelligence system for time series predictions
US6681990B2 (en) Item tracking systems and real-time inventory management
JP7402791B2 (ja) 需要予測パラメータの最適化
WO2019089146A1 (fr) Prévision de la demande à l'aide de modèles mixtes pondérés d'apprentissage automatique
US6843415B2 (en) Event-based communication in a distributed item tracking system
CN112789633A (zh) 用于机器学习模型的动态数据选择
US7519553B2 (en) Method and system for debt collection optimization
WO2019125612A1 (fr) Sélection dynamique de caractéristiques permettant la génération de modèles
EP3114620A1 (fr) Systèmes de traitement d'informations en temps réel et méthodologie basée sur un traitement homomorphique continu dans des espaces d'informations linéaires
US11854022B2 (en) Proactively predicting transaction dates based on sparse transaction data
US20200104771A1 (en) Optimized Selection of Demand Forecast Parameters
US20230094759A1 (en) System and Method of Cyclic Boosting for Explainable Supervised Machine Learning
CA3231785A1 (fr) Ajustement d'hyperparametres par ia dans une optimisation basee sur une simulation
Temizöz et al. Deep controlled learning for inventory control
Kelepouris et al. Bayesian supply chain tracking using serial-level information
US10977609B1 (en) Distribution-independent inventory approach under multiple service level targets
Xu et al. Hitting your number or not? A robust & intelligent sales forecast system
US20180268352A1 (en) Method and system for retail stock allocation
WO2003042793A2 (fr) Agent utilisant un modele predictif detaille
Li et al. Dynamic inventory allocation for seasonal merchandise at dillard’s
Lee et al. Case Fill Rate Prediction

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A2

Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY BZ CA CH CN CO CR CU CZ DE DK DM DZ EC EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX MZ NO NZ OM PH PL PT RO RU SC SD SE SG SI SK SL TJ TM TN TR TT TZ UA UG US UZ VC VN YU ZA ZM ZW

AL Designated countries for regional patents

Kind code of ref document: A2

Designated state(s): GH GM KE LS MW MZ SD SL SZ TZ UG ZM ZW AM AZ BY KG KZ MD RU TJ TM AT BE BG CH CY CZ DE DK EE ES FI FR GB GR IE IT LU MC NL PT SE SK TR BF BJ CF CG CI CM GA GN GQ GW ML MR NE SN TD TG

121 Ep: the epo has been informed by wipo that ep was designated in this application
122 Ep: pct application non-entry in european phase
NENP Non-entry into the national phase in:

Ref country code: JP

WWW Wipo information: withdrawn in national office

Country of ref document: JP