A deep real options policy for sequential service region design and timing

Abstract

There is a need for strategic flexibility in well-timed deployment decisions in the design and timing of mobility service regions, i.e. cast as ‘real options’. This problem becomes increasingly challenging considering the multiple interacting real options in such deployments. We propose a scalable machine learning (ML) based real options (RO) framework for multi-period sequential service region design and timing problem for mobility-on-demand (MoD) services, framed as a Markov decision process with non-stationary stochastic variables. A value function approximation policy from the literature uses multi-option least squares Monte Carlo simulation to get a policy value for a set of interdependent deployment decisions as deferral options (i.e., CR policy). The objective is to determine the optimal selection and timing of a set of zones to include in a service region (which to add now, which to defer to a later time to reconsider). However, prior work required explicit enumeration of all possible sequences of deployments (i.e. 5 options translate to 120 sequences). To address the combinatorial complexity arising from sequence enumeration, we propose a new variant ‘deep’ real options policy using an efficient recurrent neural network (RNN) based ML method (i.e., CR−RNN policy) to sample sequences to forego the need for enumeration, making the network design and timing policy tractable for large scale implementation. Experiments based on multiple service region scenarios in New York City demonstrate the efficacy of the proposed policy in substantially reducing the overall computational cost (i.e., time reduction associated with the RO evaluation of more than 90% of total deployment sequences is achieved), with zero to near-zero gap compared to the benchmark. We validate the model in a case study of sequential service region design for expansion of MoD services in Brooklyn, NYC, under service demand uncertainty. Results show that using the CR-RNN policy in determining optimal real options deployment strategy yields a similar performance (≈ 0.5% within the CR policy value) with significantly reduced computation time (about 5.4 times faster).