Actor-centered Agent-based Network Models

Size: px
Start display at page:

Download "Actor-centered Agent-based Network Models"

Transcription

1 Actor-centered Agent-based Network Models PI: Dr. Eduardo L. Pasiliao 1 Presenter: Dr. Tathagata Mukherjee 2 1 Munitions Directorate Air Force Research Laboratory Eglin AFB FL Intelligent Robotics, Inc. Eglin AFB FL AFOSR Dynamic Data Driven Applications Systems (DDDAS) Program Review AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

2 1 Overview of Research 2 Compound Eye Based Indoor Positioning 3 Transmitter Localization 4 Truth Finding 5 GPS Free Positioning AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

3 Research Overview Summary Effort: Design and functionality analysis of autonomous agent-based systems Each agent dynamically processes data flowing from neighbors Makes local decision/inference about the strategic construction of connections for information exchange Key Focus of Research Use network formation models from social network domain Assumption: Each actor cooperatively or non-cooperatively makes moves to optimize its objective function Decisions based on topology of local network AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

4 Learning Framework Position, time, and feature-depend Distributed machines perform dece EO/IR/RF sensor payloads. Position, time and feature dependent network topology, weights and biases Distributed machines perform decentralized decisions Decisions: when/where/how/what information transmitted Different types of sensor payloads FOR OFFIC DDDAS Actor-centered Agent- Position, time, and feature-dependent network topolo Distributed machines perform decentralized decisions EO/IR/RF sensor payloads. Distribution Statement D. Distribution authorized to the DoD and U.S. DoD contractors on AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

5 Researh Objectives Objectives: Minimize spectrum Minimize power Maximize awareness Maximize resilience Functionality: Defeat adversary Tools: Machine Learning, Deep Neural Networks, Optimization AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

6 Sensor Formation-based indoor GPSIndoor locationpositioning data. The problem is if a rob an image and solely based on the query image, ca identify what is it s location at the time of the im Image-based Indoor Localization with Compound Eyes Four Camera System Goal: Localize within building Supervised Learning Implementation: Deep Convolutional Neural Network Goal: Spatial Awareness Future: Object Detection, Tracking Implementation: Network of vision and RF sensors AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

7 Sampling Compound Image Learning Framework Figure 4: Single eye images vs compound eye images. Each single eye images and the compound eye image are of size 96x96x3. AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

8 Preliminiary Computational Results Image-based Indoor Localization with Compound Eyes Deep Convolutional Net in Matlab Training: 4 x GTX 1080 TI Error: Mean Squared Error Ground Truth: Indoor GPS system Pooling: Max Pooling Min Pooling Average Pooling Pooling type vs Accuracy Figure 6: Compound eye testing error with respect to resizing types. AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

9 Preliminary Computation Results CNN training results AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

10 Future Work on Sensor Formations Future Directions Replace 4-camera system with 8-camera payload Use truth sourcing in network formation for information exchange Use RF based sensors coupled multi-camera setup Select subset of sensors for operation based on a dynamic selection criteria Use both indoor and outdoor testbeds RF sensors based on Software Defined Radios RF Projects: Large Scale Positioning and Transmitter Localization AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

11 Transmitter Localization Problem Using single receiver find the direction of one or more transmitters Single transmitter Directional & Omni-directional 2.4 GHz Single rotating directional receiver Figure 1: Equipment Setup AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

12 Regression Technique for Direction-finding Regression Problem As bearing of transmitter is a continuous variable we use regression Support Vector Regression with ɛ-insensitive loss function Kernel Ridge Regression with squared loss Decision Tree regression ADA Boost with Decision Tree Feature engineering using ideas from Time Series data analysis Average Error: 11 ECML 2017 AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

13 Truth-Finding Problem: Find most reliable source from among several sources of information Several Sources of Information Many of them are good Commonly: Internet Uncommon: Military Applications Cooperative Control of Robotic Swarms Multi-source intelligence AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

14 Previous Work on Truth-finding Contribution: Formulation of Truth Finding as an outlier detection problem An iterative Truth Finding algorithm in high dimensional spaces using outlier removal at each step Theoretical convergence results Used algorithm for correcting records in Open Library Data Used algorithm for automatic evaluation of graduate level Algorithms midterm exam IEEE Big Data 2016 AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

15 FM Broadcast Radio-based Positioning Problem: GPS free positioning with FM signals FM broadcast band: Large & reliable coverage VHF: 88MHz-108Mhz, BW: 200kHz Less sensitive to weather condition and indoor limitation than GPS Spectrum sensing with software defined radio Large scale spectrum estimation algorithm KSJS (FM 90.5) 60dBu polygon in San Jose, CA AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

16 Previous Work on FM-based Positioning Contribution: A simple large scale localization algorithm Feature engineering for positioning Features can be used for transfer learning Theoretical results for optimality Large scale experiments on ground and aerial experiments Published Initial ACM Sigpatial 2015 Follow-up paper: Journal version with large scale experiments and transfer learning: Under revision, IEEE Transactions on Big Data Follow-up paper: Hierarchical Learning for Aerial Localization with FM, GrCon 2017 AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

17 r-centered Agent-based Network Models Actor-centered Agent-based Network Models, and feature-dependent network topology, weights and biases. chines perform decentralized decisions on when/where/how/what information is transmitted. or payloads. Statement D. Distribution authorized to the DoD and U.S. DoD contractors only (Critical Technology) (10/01/2015). Other requests shall be referred to AFRL RWWN. 18 AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18

18 Thank You AFRL/RW Actor-centered Agent-based AFOSR DDDAS / 18