October 21 – BNAIC 2004 Jonne Zutt and Cees Witteveen Multi-Agent Transport Planning Delft University of Technology
October 21 – BNAIC 2004 Contents Transportation planning Model Problem description Methods Results Future work
October 21 – BNAIC 2004 Transportation planning Guide-path design Estimating optimal number of vehicles Vehicle maintenance Order allocation Idle-vehicle positioning Vehicle routing Conflict-resolution
October 21 – BNAIC 2004 Transportation planning Guide-path design Estimating optimal number of vehicles Vehicle maintenance Order allocation Idle-vehicle positioning Vehicle routing Conflict-resolution Strategic Tactic Operational minutes hours months
October 21 – BNAIC 2004 Model Auctioneer agent Transport agent Customer agent Transport resource speed capacity max. speed capacity distance cooperative competitive
October 21 – BNAIC 2004 Model: incidents Events that disrupt regular plan execution and generally require re-planning Examples: customers that change or retract transportation orders, unpredictable congestion, vehicle break-down, communication failure Incidents are generated proportional to the resources. Pfail = 0.x means each resources is expected to fail x·10% of the time.
October 21 – BNAIC 2004 Problem description Find conflict-free routes for the operational agents such as to execute all orders, to maximize rewards and to minimize costs Maintain feasible plans even when incidents occur
October 21 – BNAIC 2004 Conflicts 1.Resources have limited capacity ABC 2.Instantaneous exchange ABD Time ABC AB D
October 21 – BNAIC 2004 Method Do (partial) order assignment While agents are not ready –Compute traffic-aware shortest path –Agent compete who schedules first (P1) –Winner schedules n resources (P2) If current order rewards are below threshold, agent tries to reroute (P3)
October 21 – BNAIC 2004 Method: traffic-aware shortest path Agents know which time-windows are in use by other agents per resource Run an A* algorithm: store routes on open list, check for conflict when appending to candidate route Process is guaranteed to terminate and find the traffic-aware shortest path
October 21 – BNAIC 2004 Method: agent selection functions (P1) Random Provides a baseline for the others Delays Agent with maximum wait time first Deadlines Agent with most strict deadlines first Penalties Agent with lowest planned reward first
October 21 – BNAIC 2004 Method: resource block-size (P2) How many resources (fraction of route) are scheduled after the agent is selected by the agent selection function? Hypothesis: a small block-size slightly increases performance but also increases computation time
October 21 – BNAIC 2004 Experiments 10 transport networks with 25 resources 10 sets of transportation orders with 75 random orders each 2 different sets of agents with 30 randomly located agents each Incidents with failure probability 0, 0.1 and 0.2.
October 21 – BNAIC 2004 Agent selection Average sum of delivery penalties No incidents Pfail = 0.1Pfail = reroutes1 reroute 0 reroutes1 reroute0 reroutes1 reroute 1.Random 2.Delays 3.Deadlines 4.Penalties
October 21 – BNAIC 2004 Block size No incidents Pfail = 0.1Pfail = Average sum of delivery penalties 2246∞24∞2∞2∞6∞246∞ 1.max. number of reroutes 2.block size
October 21 – BNAIC 2004 Time for different block sizes No incidents Pfail = 0.1Pfail = ∞24∞2∞2∞6∞246∞ Average cpu time max. number of reroutes 2.block size
October 21 – BNAIC 2004 Conclusions Rerouting instead of static routing improves performance Prioritizing agents according to the order’s deadlines seems to work well (on random network topologies) Smaller block size is better, at the cost of some extra computation time
October 21 – BNAIC 2004 Future work Experiments for different network topologies (Market-based) distributed version Effect of coordination strategies
October 21 – BNAIC 2004 Questions CABS project: My homepage: My