Decision Support for Attended Home Deliveries in Metropolitan Areas

Transcription

1 Platzhalter für Bild, Bild auf Titelfolie hinter das Logo einsetzen Decision Support for Attended Home Deliveries in Metropolitan Areas Ann Melissa Campbell, Jan Fabian Ehmke 2013 Service Management and Science Forum

2 Challenges for Attended Home Deliveries Online retailing is the fastest growing retail sector Efficient and reliable delivery of orders is critical for lasting success This is particularly challenging for attended deliveries Examples: grocery, large appliances, repairmen

3 Challenges for Attended Home Deliveries For attended deliveries Customers and service providers agree on a time window Customers expect on-time deliveries, narrow time windows Online retailers need to maximize the number of customers visited Deliveries in metropolitan areas suffer from congestion Not surprising many of the early businesses failed!

4 PTV, Karlsruhe Key Decisions Strategic: Decide what service time windows to offer in each customer area. Operational: 1. Selection of a service time window Decide what windows are feasible for each customer Decide when to close time windows in different areas 2. Order picking Picking in warehouse vs. stores 3. Vehicle routing and scheduling Minimizing costs of delivery

5 PTV, Karlsruhe Related Work Strategic: Decide what service time windows to offer in each customer area. Agatz, Campbell, Fleischmann, Savelsbergh 2011 Operational: 1. Selection of a service time window Assumed deterministic travel time Not look at impact of congestion/time dependency Campbell & Savelsbergh 2005/2006, Asdemir et al Order picking 3. Vehicle routing and scheduling Consider a time-dependent cost matrix TDVRPTW Ichoua et al. 2003, Eglese et al. 2006, Fleischmann et al. 2004, Maden et al. 2010

6 Our Focus How do we decide which requests to accept in each service time window? How do we balance the need to provide reliable service with profit maximization? Examine several acceptance mechanisms Differ in level of information, ease of implementation Evaluate them using simulation framework Simulated demands Simulated travel times to reflect congestion in metropolitan areas

7 Home Delivery Problem Retailer offers a predefined set of time slots on a day Delivery requests arrive before start of service Customer selects time slot Each customer has first and second choice Provide first choice if available Provide second choice if first choice not available Customer leaves if neither available Need to quickly decide if each request can be handled in a particular time slot Objective: maximize the number of accepted requests Assumption: vehicle capacity not binding Time Slot [12:00,12:59] [13:00,13:59] [14:00,14:59] [15:00,15:59] [16:00,16:59] [17:00,17:59] [18:00,18:59]

8 Solving the Home Delivery Problem Need to quickly decide if each request can be handled in a particular time slot This can be done in a rough/approximate way Rules of thumb Fast/easy Refer to as static approaches This can be done in a more dynamic way Use more detailed information, build routes Can accept more deliveries May accept too many deliveries! Congestion, stochastic travel times Refer to as dynamic approaches Time Slot [12:00,12:59] [13:00,13:59] [14:00,14:59] [15:00,15:59] [16:00,16:59] [17:00,17:59] [18:00,18:59]

9 Home Delivery Problem Static Approaches SLOT: Accept a fixed number of requests per time slot If first priority time slot cannot be accommodated, check alternative option, else: reject Common/simplest Time Slot SLOT TD-SLOT [12:00,12:59] 6 7 [13:00,13:59] 6 7 [14:00,14:59] 6 6 [15:00,15:59] 6 5 [16:00,16:59] 6 4 TD-SLOT: Allow the number of requests to vary depending on time of day Requires general knowledge on the evolution of travel times [17:00,17:59] 6 4 [18:00,18:59] 6 5 Feasibility of scheduled routes is not guaranteed.

10 Home Delivery Problem Dynamic Approaches (1/2) Dynamic (DYN): Incorporate time-dependent travel times and precise routing Consider a complete TDVRPTW solution for feasibility check ( I1/Solomon 1987) If the current request (and the already accepted requests) can be accommodated feasibly, it is accepted This can still fail due to stochastic travel times! Dynamic Simple Buffer (DYN-SBF): Include a simple buffer time bt in the feasibility check of each request Accept a request i only if arr i + bt < late i E.g. get there 10 minutes before window closes All customers have the same buffer

11 Home Delivery Problem Dynamic Approaches (2/2) Dynamic Buffer (DYN-BUF): May need more dynamic buffer bt i Does not treat all requests the same Observation: a long drive yields more variability in the arrival time Buffer size should change based on travel time Observation: Think of last bus stop. The bus rarely gets there on time The later in the tour, more buffer needed due to accumulated lateness Buffers change based on location in the tour Idea: accept a request i only if arr i + bt i < late i Computing this buffer correctly is non-trivial!

12 Home Delivery Problem Propagation of Arrival Times (1/2) Arrival time distribution at customer not based on simply combining means and variances of travel times Many assume this! Arrival time at one customer based on previous travel times and time windows arr i arr i earl i late i (a) (b) Consider if majority of distribution of arrival times is before a customer s time window The opening of the window determines the departure time This reduces the variance in arrival times at subsequent customers Summary: Service time windows impact how you propagate the variability earl i earl i arr i late i late i (c)

13 Home Delivery Problem Propagation of Arrival Times (2/2) Our idea: A recursive calculation of variance that depends on: Proportion β i of arrival time distribution after earl i 2 Variation of travel times σ i 1,i arr i arr i earl i late i (a) (b) γ i = cumulated standard deviation of arrival time earl i late i γ i = 2 2 β i 1 γ i 1 + σ i 1,i, γ 0, β 0 = 0 bt i = α β i γ i With user specific service level α arr i (c) earl i late i Summary: We compute a unique buffer for each customer based on information from earlier in the tour

14 Impact of DYN-BUF TW 15:00-16:00 DYN: Arrival 15:56 BUF: Arrival 15:32 TW 15:00-16:00 DYN: Arrival 15:59 BUF: Arrival 15:51 Platzhalter für Bild, Bild auf Titelfolie hinter das Logo einsetzen TW 18:00-19:00 DYN: Arrival 18:58 BUF: Arrival 18:14 TW 17:00-18:00 DYN: Arrival 17:46 BUF: Arrival 17:35 DYN: 34 customers Proportion served late: 94% BUF: 34 customers Proportion served late: 7% TW 19:00-20:00 DYN: Arrival 19:42 BUF: Arrival 19:28

15 Computational Experiments Experimental Design Real-world inspired metropolitan network Two zones of customer locations Inner city and suburbs 100 instances from each zone Mean time-dependent travel times generated from speed multipliers To reflect daily congestion patterns Parameters of distribution based on fit with real travel time data (Stuttgart) Basic test: 7 time slots of width: 60 minutes Service time at customers: 20 minutes Three vehicles Solve each instance with each acceptance mechanism Simulate 1000 times cluster 1 cluster 2 cluster 3 cluster 4 cluster 5 cluster 6

16 Computational Experiments Suburban Delivery SLOT accepts the least requests DYN accepts the most but is late in 86% of realizations DYN-BUF: smaller request acceptance, but occurrence of lateness is reduced a lot DYN-SBF and static approaches accept far fewer but are also rarely late tend to be conservative 100% 98% 96% 94% 92% 90% 88% 86% 84% 82% 80% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% avg # of accepted requests (suburban) SLOT TD-SLOT DYN-SBF DYN-BUF(2.0) DYN avg prop. of tour plans late (suburban) SLOT TD-SLOT DYN-SBF DYN-BUF(2.0) DYN

17 Computational Experiments Inner City Delivery Extent of lateness is less due to shorter distances DYN-SBF is very strict in inner city areas fewer deliveries than TD-SLOT DYN-BUF reduces lateness occurrence, but accepts almost as many requests as DYN 100% 98% 96% 94% 92% 90% 88% 86% 84% 82% 80% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% avg # of accepted requests (inner city) SLOT TD-SLOT DYN-SBF DYN-BUF(2.0) DYN avg prop. of tour plans late (inner city) SLOT TD-SLOT DYN-SBF DYN-BUF(2.0) DYN

18 Conclusions and Outlook Static approaches do not maximize utilization of logistics resources Parameter estimation difficult and inflexible Dynamic approaches are able to increase efficiency Basic dynamic approaches good for close locations in downtown Most important to consider travel time variation with longer distances + congestion (suburbs) shorter time windows Design routing algorithm to minimize lateness propagation

19 Platzhalter für Bild, Bild auf Titelfolie hinter das Logo einsetzen Thank you for your attention!