Towards Dynamic Tracking of Multi-Agents Teams An Initial

Towards Dynamic Tracking of Multi-Agents Teams: An Initial Report Dorit Avrahami-Zilberbrand Gal A. Kaminka The MAVERICK Group Computer Science Department Bar Ilan University, Israel galk@cs. biu. ac. il avrahad 1@cs. biu. ac. il

Plan Recognition for Dynamic Multi-Agents Teams n Single-agent plan recognition n n Inferring the intentions, plans, goals of an agent Based on observations of its interactions with environment Examples: NLP, intrusion detection systems, surveillance Team plan recognition n Tracks dynamic grouping and ungrouping of agents Inferring team’s joint goals and plans Identifying interactions between groups of agents avrahad 1@cs. biu. ac. il 2

Background n Utilize pre-defined information to detect groups n YOYO (Kaminka & Bowling 2002): Static social structure n RESC-team (Tambe 1996): Plans expected to be agreed-upon n STABR (Sukthankar & Sycara 2006): Pre-specified geometric formation n n Intille & Bobick 1999: Constraints among football players Hongeng & Nevatia 2001: Temporal constraints between agents Cannot detect grouping and ungrouping of teams Our goal: n Detect dynamic splitting and merging of groups n No reliance on static information about groups avrahad 1@cs. biu. ac. il 3

Queue cutting problem n n Two friends stand in line (Andy & Bandy) Andy leaves to go to the bathroom Bandy makes progress with the line When coming back, Andy joins Bandy. Requires tracking dynamic joins and splits avrahad 1@cs. biu. ac. il 4

Recognizing Dynamic Multi-Agent Groups n Dynamic data structure n n n Utilize information from group of agents n n Dynamic Hierarchical Group Model (DHGM) Tracks dynamic grouping and ungrouping of agents Identify the interactions between groups of agents Identify potential suspicious behavior Understand better agent’s actions (queue-cutting problem) Efficient: DHGM encapsulates shared data of groups avrahad 1@cs. biu. ac. il 5

Background: Hierarchical Plan Library n n n Directed acyclic connected graph Vertices denote plan steps Edges n n Vertical (decomposition) edges n Horizontal (sequential) edges security position X-Ray coffee Self cycles model durations Each plan step contains conditions Plan path: Root to leaf path avrahad 1@cs. biu. ac. il 6

The Symbolic Recognizer (Avrahami-Zilberbrand Kaminka IJCAI, MOO 2005) n n Input: vector of observed features Efficiently matches observations to plan steps n n Tagging and propagating n n n Feature Decision Tree Tags with observation time-stamp and propagates Output: paths tagged with time-stamp t Advantages n n Handles key capabilities in plan recognition Efficient – linear time in the plan library size avrahad 1@cs. biu. ac. il 7

Symbolic Recognizer - Example 1 2 root 2 1 1 2 security entrance 1 position coffee 1 position 2 X-Ray Shop 2 without bag 2 with bag 1 position 1 board 2 X-Ray 2 without bag avrahad 1@cs. biu. ac. il 1 position coffee gate 2 with bag 8

DHGM: Dynamic Hierarchy Group Model n Dynamically-maintained structure n n Updated with every observation Reflects the current groups Reflects the history of these groups Each node in DHGM n n Represents a group: Agents that are executing the same behavior Points to recognized leaves in the plan library Keeps track of how long the group was in this branch From each node there are branches to sub-groups avrahad 1@cs. biu. ac. il 9

DHGM maintenance process Bottom-up traversal of DHGM; for each node N: n If N leaf: n n Create new temporary agent array For each agent, execute SBR on plan-library for N n n Update temporary agent array to point to new results Create subgroup branches if needed, update time-stamp Update agent array with new temporary array If N non-leaf n n Merge sub-group branches that point to same plan-steps Update time-stamp avrahad 1@cs. biu. ac. il 10

Example: Track 100 agents on position x-ray gate t=1 t=2 100 Position t=3 t=4 100 50 50 25 25 With Bag Without Bag 50 20 100 Position 50 30 With Bag Without 50 Bag gate 25 gate avrahad 1@cs. biu. ac. il 25 24 1 Position 11

n n Space Complexity Analysis Complexity O(Lg) n L is the plan library size n g is the maximum number of groups n Compared to. O(Ln) running individual (n is number of agents) Run-time complexity of single update O(n. LD + nlogm) n L is the plan library size n is the number of agents n D is the depth of the plan library n m is the number of agents in a group n Compare to naïve O((n. LD)2) = O(n 2 L 2 D 2) n avrahad 1@cs. biu. ac. il 12

Detecting Suspicious Behavior using DHGM n n Suspicious behavior recognition for individual tracking n Explicit Recognition: Plan library represents suspicious behavior n Implicit Recognition: Plan library represents normal behavior The challenge: n n Suspicious behaviors that can not be captured by individual tracking The Solution: n Utilize the DHGM for detecting suspects (explicitly) avrahad 1@cs. biu. ac. il 13

Heuristics for detecting suspicious behavior: Focusing on an agent Agent behave differently from its group is a suspect n n Agent did not return to his group k time-stamps Rest of the group is m times bigger from suspected group Example: Passenger that did not stand in security line avrahad 1@cs. biu. ac. il 14

Heuristics for detecting suspicious behavior: Example t=4 100 50 50 gate 25 24 1 Position avrahad 1@cs. biu. ac. il 15

Heuristics for detecting suspicious behavior: Clearing an agent n n Clear an agent n If it behaves normally wrt its group history n Example: queue-cutting problem Using DHGM n n n Individual P. R. returns no results after propagating phase Compare SBR matching phase results to agent’s history If executing same plan-steps as its group: Clear avrahad 1@cs. biu. ac. il 16

Summary and Future work n This is initial work n n n Combination of single-agent plan recognizer and the DHGM Two heuristics for recognizing suspicious behavior Future work n n Formally define the recognition queries Closely examine run-time tradeoffs Come see our AAAI talk (Wed 10: 20): Incorporating observer-biases in plan recognition avrahad 1@cs. biu. ac. il 17