BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Seminar by Prof. Stefan Feuerriegel\, ETHZ
DTSTART:20190620T150000
DTEND:20190620T163000
DTSTAMP:20260509T235638Z
UID:6a5bba18013b53699052ce0a1438368d6541036cd0ae234ed7c112d1
CATEGORIES:Conferences - Seminars
DESCRIPTION:Séminaire par Prof. Stefan Feuerriegel\, ETHZ\n"Effective Pol
 ice Patrolling: A Stratified Monte Carlo Tree Search"\n\nRésumé\nPolice 
 patrolling was proven to prevent crime\, and yet a corresponding decision
  problem that aids police management is lacking.  As a remedy\, this wor
 k formalizes effective police patrolling as a finite-horizon discrete-tim
 e Markov decision process (MDP). It yields a policy for choosing optimal 
 pathways for multiple patrols with the objective of reducing crime risk. 
 The MDP for multiple patrols yields a high-dimensional state-action space
 \, which prohibits straightforward solutions. Therefore\, we apply Monte 
 Carlo tree search (MCTS) as a solution technique and\, in order to obtain
  substantially better performance\, develop a stratified MCTS: that is\, 
 we translate the discrete MDP into sub-problems of continuous optimal pat
 h planning modeled via the eikonal equation\, i.e.\, a non-linear partial
  differential equation\, where the optimal pathways correspond to soluti
 ons of ordinary differential equations. Given this formalization\, we der
 ive two novel MCTS roll-out policies\, namely\, a Dijkstra policy and a f
 ast-marching policy. This allows us to provide theoretical guarantees tha
 t\, under mild assumptions\, both roll-out policies choose pathways that 
 are locally optimal. The superiority of our stratified MCTS over a naive 
 MCTS is confirmed in an extensive series of experiments. Furthermore\, th
 e decision heuristics which are currently applied in police practice are 
 considerably outperformed. Our work entails important implications. Publi
 c sector management benefits from an effective use of resources and\, met
 hodologically\, we establish that tailored\, domain-specific roll-out pol
 icies can achieve theoretical optimality guarantees.
LOCATION:ODY 4 03 https://plan.epfl.ch/?room==ODY%204%2003
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR