International Business Machines Corporation

Shivaram Subramanian

Pavithra Harsha

Wei Sun

Markus Ettl

Brian Leo Quanz

Ramesh Natarajan

Soumyadip Ghosh

Xiaoxuan Zhang

Max Biggs

Jonathan R Hosking

Jonathan R M Hosking

Rajesh Kumar Ravi

A machine learning system that incorporates arbitrary constraints is provided. The machine learning system selects a set of domain-specific constraints from a plurality of sets of domain-specific constraints. The machine learning system selects a set of general functional relationships from a plurality of sets of general functional relationships. The machine learning system maps the selected set of general functional relationships and the selected set of domain-specific constraints to a set of learning transforms. The machine learning system modifies a machine learning specification according to the set of learning transforms, wherein the machine learning specification specifies a model construction, a model setup, and a training objective function. The machine learning system optimizes a machine learning model according to the modified machine learning specification.

Domain-specific constraints for predictive modeling

A machine learning system that incorporates arbitrary constraints into deep learning model is provided. The machine learning system provides a set of penalty data points en a set of arbitrary constraints in addition to a set of original training data points. The machine learning system assigns a penalty to each penalty data point in the set of penalty data points. The machine learning system optimizes a machine learning model by solving an objective function based on an original loss function and a penalty loss function. The original loss function is evaluated over a set of original training data points and the penalty loss function is evaluated over the set of penalty data points. The machine learning system provides the optimized machine learning model based on a solution of the objective function.

Loss augmentation for predictive modeling

A machine learning system that uses a split net configuration to incorporate arbitrary constraints receives a set of input data and a set of functional constraints. The machine learning system jointly optimizes a deep learning model by using the set of input data and a wide learning model by using the set of constraints. The deep learning model includes an input layer, an output layer, and an intermediate layer between the input layer and the output layer. The wide learning model includes an input layer and an output layer but no intermediate layer. The machine learning system provides a machine learning model comprising the optimized deep learning model and the optimized wide learning model.

Split-net configuration for predictive modeling

A computer generates an optimized decision distribution vector for a plurality of related, demand-correlated products. The computer receives data indexed by product, with each entry including several entry attributes. The computer receives decision context data for the products. The computer determines a set of primary attributes and trains a first machine learning model based upon those attributes. The computer receives a decision optimization request that includes an associated set of attributes corresponding to the primary attributes. The computer scores the associated set of attributes, using the first machine learning model, to generate a baseline purchase propensity. The computer trains a second machine learning model, based upon the baseline purchase propensity and the decision context data, to generate own-product and cross-product elasticity data. The computer, using the own-product and cross-product elasticity data, generates a decision distribution vector for the group of related, demand-correlated products.

Automated hybrid propensity decision vector generation using artificial intelligence

A system that compresses data during neural network training. A memory stores computer executable components and neural network data, and a processor executes computer executable components stored in the memory. An anticipatory value of inventory (VOI) optimization component calculates optimal VOI and prices for immediate-future inventory levels in parallel and writes latest price updates for respective states to a price stack; and a recommendation component provides customized pricing recommendation for a product relative to a unique customer as a function of the latest price updates for respective states to the price stack.

Real time personalized pricing for limited inventory assortments in a high-volume business environment

A predictive-control approach allows an electricity provider to monitor and proactively manage peak and off-peak residential intra-day electricity usage in an emerging smart energy grid using time-dependent dynamic pricing incentives. The daily load is modeled as time-shifted, but cost-differentiated and substitutable, copies of the continuously-consumed electricity resource, and a consumer-choice prediction model is constructed to forecast the corresponding intra-day shares of total daily load according to this model. This is embedded within an optimization framework for managing the daily electricity usage. A series of transformations are employed, including the reformulation-linearization technique (RLT) to obtain a Mixed-Integer Programming (MIP) model representation of the resulting nonlinear optimization problem. In addition, various regulatory and pricing constraints are incorporated in conjunction with the specified profit and capacity utilization objectives.

Managing time-substitutable electricity usage using dynamic controls

System and method for omni-channel retailer operations that integrate a network of brick-and-mortar stores with their online channel. The system and method includes calibrating a demand model for both brick-and-mortar sales and on-line channels over which a product is sold, the calibrating based upon a cross-channel fulfillment-aware inventory effect. An omni-channel sales prediction and fulfillment model is then constructed based on the calibrated demand model. Using constructed linear demand and revenue models, a plan is generated to optimize one or more: allocating of the product across physical stores, partitioning of the product for sales, and pricing of the product. Customer choices are jointly modeled across channels to allow switching, and a ship-from-store (SFS) inventory effect feature only for brick choice is applied to capture asymmetry. Inventory decision variables are introduced into an omni-channel pricing formulation to manage SFS-effect induced non-convexity and specific reformulations applied to recover a linear MIP.

Managing cross-channel fulfillment impact within shared inventory demand systems

Systems and methods for modeling and forecasting cyclical demand systems in the presence of dynamic control or dynamic incentives. A method for modeling a cyclical demand system comprises obtaining historical data on one or more demand measurements over a plurality of demand cycles, obtaining historical data on incentive signals over the plurality of demand cycles, constructing a model using the obtained historical data on the one or more demand measurements and the incentive signals, wherein constructing the model comprises specifying a state-space model, specifying variance parameters in the model, and estimating unknown variance parameters.

System and method for modeling and forecasting cyclical demand systems with dynamic controls and dynamic incentives

A computer implemented method and system of allocating a target commodity product onto an omnichannel distribution system is provided. Historical data related to the target commodity product is retrieved. Data mining is performed on the retrieved historical data to identify patterns therefrom. An omnichannel nominal demand prediction model is developed based on the identified patterns of the data mining for the omnichannel distribution system. A life-cycle demand is forecast. An allocation for the target commodity product based on the omnichannel nominal demand prediction model is created. A worst case scenario of allocation of the target commodity product for the omnichannel distribution system is identified. The allocation is adjusted to prevent the worst case scenario.

Allocating a product inventory to an omnichannel distribution supply chain

A streams platform is used. Multiple streams of electricity usage data are received, each from an electrical meter providing periodic updates to electrical usage for devices connected to the electrical meter. Weather information is received corresponding to locations where the electrical meters are. Real-time predictive modeling of electricity demand is performed based on the received multiple streams of electricity usage data and the received weather information, at least by performing: updating a state space model for electrical load curves using the usage data from the streams and the weather, wherein the updating uses current load observations for the multiple streams for a current time period; and creating forecast(s) for the electricity demand. The forecast(s) of the electricity demand are output. Appliance-level predictions may be made and used, and substitution effects and load management functions may be performed.

Real-time forecasting of electricity demand in a streams-based architecture with applications

Embodiments are directed to a computer implemented method of generating inventory valuation data for an omnichannel (OC) retail operation. The method starts with an unsolvable OC nonlinear nonconvex problem, applies transformations to generate a mixed-integer program (MIP) that is a tractable linear nonconvex form, solves the MIP, fixes prices at optimal values to achieve dimensionality reduction and eliminate all non-convexity by eliminating the pricing dimension. The method further obtains inventory flow linear programming (LP) that is linear convex, and solves the LP to recover a dual solution as initial inventory valuations.

Growing community of inventors

New Fairfield, CT, United States of America

Shivaram Subramanian