US 12,282,949 B2
System and method for forecasting real estate solutions
Marshall O'Moore, New York, NY (US); Ruiqi Hu, Riverside, CT (US); Christelle Bron, New York, NY (US); and Derek Chanler-Berat, Bedford Corners, NY (US)
Assigned to CBRE, INC., Dallas, TX (US)
Filed by CBRE, INC., Dallas, TX (US)
Filed on Feb. 5, 2024, as Appl. No. 18/432,760.
Application 18/432,760 is a continuation of application No. 17/076,374, filed on Oct. 21, 2020, granted, now 11,893,625.
Claims priority of provisional application 62/924,846, filed on Oct. 23, 2019.
Prior Publication US 2024/0177228 A1, May 30, 2024
This patent is subject to a terminal disclaimer.
Int. Cl. G06Q 30/0645 (2023.01); G06N 3/08 (2023.01); G06N 5/046 (2023.01); G06Q 10/10 (2023.01); G06Q 50/16 (2012.01)
CPC G06Q 30/0645 (2013.01) [G06N 3/08 (2013.01); G06N 5/046 (2013.01); G06Q 10/10 (2013.01); G06Q 50/16 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method for optimizing future real estate solutions, comprising:
receiving, by a computing device, historical data including a historical headcount associated with an entity;
generating, by the computing device, a forecast of a headcount demand for the entity at multiple future time points based on the historical data;
creating, by the computing device, a plurality of real estate scenarios based on the forecast of the headcount demand;
constructing, by the computing device, one or more option trees corresponding to the plurality of real estate scenarios, each option tree comprising a set of decision points and associated real estate actions;
initializing, by the computing device, a policy network with a randomly weighted policy function and a value network with a randomly weighted value function;
iteratively training, by the computing device, the value network to predict future costs associated with the real estate actions by processing inputs from the option trees and minimizing a predictive error using an optimization algorithm;
iteratively training, by the computing device, the policy network to recommend real estate actions that minimize predicted future costs based on outputs from the value network; and
alternating, by the computing device, the training of the policy network and the value network until a convergence criterion is met, wherein the convergence criterion is based on a stabilization of network weights or a minimization in prediction error below a predetermined threshold.