World Model implementation & results Workshop Stuttgart, 5/6 November 2009 Rob Janssen
Why World Model? Before WM: - knowledge acquired locally (little communication) - no consensus on environment - no team play (only role-tasks) With WM: - share knowledge of environment - agree on own/opponent/ball positions - allows for passing, avoidance etc..
Features World Model Cluster shared information Provide consensus Run in real-time But.. Reliable communication necessary (multicast packet loss, delay)
Local World Model Omnivision (segment black blobs)
Local World Model Clustering blobs (Schubert & Sidenbladh 2005) For each report: -Generate hypotheses -Propagate objects in hypotheses (linear Kalman) -Update likelihood (compare report with objects) -Prune hypotheses tree -MAPE (obtain hypothesis with highest probability)
Local World Model Local opponents send to World Model
Local/Global World Model
Global World Model Same as local world model Except: -Labeling added -Label consistency with switching hypotheses -Agent ID overrules label -Agent position overrules clustered object position
Result WK Graz Cambada 2nd half
Validation 1) Topcamera above field 2) Match 3vs3
Future work - Ball: Non-linear observer not yet implemented Locally obtained/communicated -Strategic learning -Improve communication!!
Questions