Eisenstein and Iyer [41] approach this problem by devising
flexible schedules for garbage trucks in the city of Chicago. The
weight and time required to collect garbage from a single block are
modeled as random variables, whose parameters are estimated
from real data. They introduce a dynamic scheduling scheme that
uses flexible truck routes, with the flexibility deriving from the
possibility to visit the dumpsite once or twice a day. The choice of
a truck route is modeled as a Markov Decision Process (MDP), that
adjusts the number of dumpsite visits throughout the week to
maximize the service level. Given a sequence of n blocks to be
collected by a single truck over a work week, plus an additional
dummy block nþ1 with no weight and time requirements, the
state of the MDP is said to be i at day t if at the beginning of day t
the truck is starting to collect block i. At the beginning of each
morning, a decision a must be taken (aAA ¼ f1; 2g), concerning
whether to use a route which visits the dumpsite once (a¼1), or
twice a day (a¼2). Given the probabilities pðjji; aÞ to transition to
state j following decision a in state i, the value function vt(i) is
defined as the maximum probability that a truck will complete all
assigned blocks by the last day of the week, given that it starts in
state i on day t. Since the system always begins in state 1 on day 1,
the objective is to determine a policy, i.e., a sequence of deterministic
Markov decision rules, that maximizes v1ð1Þ. This is achieved
through backward induction. Empirical results of the impact on
data collected from sample wards show the possibility to obtain a
reduction in the number of truck routes required to perform the
garbage collection operations, ranging between 12% and 16%,
leading to over $9 million overall cost saving