Reporting Down Under A CNAS update Daniel D. Corkill Department of Computer Science University of Massachusetts Amherst Kevin Bartlett, Robert Fleishauer,

Reporting Down UnderA CNAS update

Daniel D. CorkillDepartment of Computer Science

University of Massachusetts Amherst

•Kevin Bartlett, Robert Fleishauer, Walter Koziarz, Wenchian Lee, Zenon Pyke, Wilmar Sifre, Lok Kwong Yan, and Paul Yaworski of AFRL and Huzaifa Zafar of UMass contributed to CNAS. Douglas Holzhauer (now retired from AFRL and teaching at SUNYIT) initiated the CNAS project.

Huzaifa Zafar

Atmospheric Monitoring

•Ground-level environmental monitoring•increased air-drop precision &all-weather landing safety

•forest fire detection, monitoring•airborne hazardous materials (hazmat incidents)

•Battery-powered, ad hoc sensor networks•can provide accurate local atmospheric conditions

•Long-term objective•air-dropped, self-organizing sensor agents• potential for undesirable geographical/meteorological placement

CNAS (Cognitive/Collaborative Network for Atmospheric Sensing)

•Experimental, agent-based, power-aware sensor network for ground-level atmospheric monitoring•research & demonstration tool•realistic exploration of agent-based environmental monitoring

•blackboard-system & MAStechnologies (CLISP & GBBopen)

•hardware capabilities that are likely to become cost-effective for production deployment in the next few years

Project Objective

•Develop a cognitive, collaborating-software framework for resource-aware sensor agents

• Sensor agents work together to achieve overall goals•sensor agents are part of a larger organization•may need to satisfy more global goals at the expense of local goals

• Agile, responsive, opportunistic behavior•adaptable, purposeful, graceful degradation

• Meta-level awareness/reasoning of resource use•sensors, processing, communication, power

• Organizational agents•understand and respond beyond their local situation

• Support rapid inclusion & customization of new methods in fielded sensor networks

Sensor-Agent Characteristics

•Sparse deployment•sensor agents are located near the limitof wireless communication range•few alternate message routes•most communication is multi-hop

•Each agent’s WiFi adapter must be turned off as much as possible•even listening consumes significant power

•Sensor agents obtain and process local-environment readings once every second•even when WiFi is powered off

Sensor-Agent Hardware

•ISI’s PASTA microsensor platform(Power-Aware Sensing, Tracking, & Analysis)• Intel PXA255-based CPU• distributed-peer component architecture • 64MB non-paged system memory• compact-flash supplemental “disk”

•Crossbow MTS420CA• Intersema MS5534AM barometric pressure• Sensirion SHT11 relative humidity/temperature• TAOS TSL2550D ambient light• Leadtek GPS-9546 GPS module• Analog Devices ADXL202JE dual-axis accelerometer

•Netgear MA111 USB wireless adapter

TACMET-Augmented Sensor Agent

•Basic sensor agent + Climatronics TACMET II 102254 •temperature sensor •fast-response, capacitive, relative humidity sensor

•barometric pressure sensor•flux gate compass•folded-path, low-power sonic anemometer

•USAF certified•provides cross-correlation with Crossbow

Additional Nodes & Roles

•Sensor agents acting as cluster head nodes•aggregate and maintain summariesof readings taken over geographic regions of interests

•Mobile console nodes mayenter and leave the CNAS monitoring area•Obtaining regional summaries and individual sensor-agent data

•Can change CNAS tasking and policies

•Regional node connected to the internet

Sensing Activities

•Summary readings are computed from the 1-second readings every 5 minutes:

• temperature (5-minute average)• dew point (5-minute average)• pressure (last reading)• altimeter (last reading)• wind-u-component (2-minute average, TACMET agent)• wind-v-component (2-minute average, TACMET agent)

•Individual summaries are retained at the agent & sent to the cluster head sensor agent

• also sent to regional node/console nodes hourly

•Cluster head computes 5-minute cluster averages

•Regional node generates METAR reports from cluster averages

• one report every 15 minutes

Cluster Head

•Role assumed by any sensor agent in the cluster•cluster consists of all sensor agents located in a non-overlapping geographic region of interest

•defined by meteorologists; known by all agents

•add’l responsibilities to regional & cluster nodes

•total preference order for role assignment•based on sensor agent locations

•Reassigned dynamically•current reachability status

•determined locally

•cluster can become bifurcated•recombined when connectivity re-established

cluster head

Power Expenditure & Communication

•Sensor agent•12V battery12,000mA-hours power

•WiFi adaptor•largest power expenditure ratewhen powered on

•cannot turn off unilaterally•others may want to contact agent•may be acting as forwarders in multi-hop transmissions

•all agents turn on/off radios at the same time

•set of compatible time-window policies•selected adaptively based on anticipated need

Power Expenditure & Communication

•Many communication & routing protocols have been developed for wireless sensor networks•Akkaya and Younis provide a recent survey•most assume stationary sensors (like CNAS)•some assume mobile sinks (also like CNAS) & periodic reporting requirements

•most focus on transmission distance/quality as cost & assume full-time listeners

•others, such as GAF, switch off a percentage of nodes in an area—enough remain on for forwarding

•Oft-assumed characteristics are not present with CNAS

Communication Window

• Time aligned policies:• hourly: top of each hour• half-hourly: every 30 minutes• quarter-hourly: every 15 minutes• hourly-overnight-sleep: hourly, but no communication 6PM-6AM

• half-hourly-overnight-sleep: …• quarter-hourly-overnight-sleep: …

Communication Window

•Sensor agent activities continue throughout radio-off periods•Communications to/from sensor agents buffered until next window

•Delivery failure also buffers messages; redelivery attempted when communication restored

•Time-aligned policies facilitate resynchronization to changed policy

•WiFi shutdown can be temporarily overridden from the regional node to facilitate debugging

•but not until the next communication window…!

Initial CNAS deployment: PATRIOT 2006•CNAS’s first field test

•2006 PATRIOT Joint Exercise (Fort McCoy, Wisconsin)

• 9 “basic” sensor agents & 8 TACMET-augmented sensor agents

• As reported here, last year: successful, but…•transient & permanent sensor-agent failures (heat)

•lack of Crossbow & external-communication

•Queensland, Australia (June—July 2007)•Two deployments

•Drop zone (similar to PATRIOT 2006)•Actually deployed twice; once in 15 minutes

•Taller native vegetation issue• solved with broom handles

•Regional-node (laptop) failure

•UTOF: Congested (urban) setting •Brick, concrete, & steel walls•Periodic radio interference fromnearby system

• CNAS recovered gracefully, without intervention

Talisman-Saber Combined Exercise 2007

•Complete technical success

•posting of weather observations to AFWA and BOM weather servers

•support of COUNTER small UAVs

•support of airdrop operations

•adapting to changing requirements during Exercise

•Invited to participate in Talisman-Saber 2009 Exercise

Talisman-Saber Combined Exercise 2007

Latest work: Solar harvesting

•Solar panel at each sensor agent•battery reserves can grow(up to full capacity)

•Replenishment prediction•sunshine forecast available from outside sources•activity decisions take expected replenishment into account

•each agent must learn its solar visibility/shading•indirect received-energy indication via battery voltage

•agents interact with nearby agents to discover what times of day they are shaded

•will be described in the next presentation…

Latest work: Silence with responsiveness•Persistent routing tables

•eliminate OLSR re-initialization at the start of each communication window

•assume no change has occurred during radio off•equivalent to all changes occurring as a burst at window start

•application-level communication is available immediately•routing adaptation occurs concurrently with application messages

•evaluated on a small indoor CNAS network (Zenon Pyke)

•routing stabilized quickly (a few secs versus 40-50 to re-init)

•sparse outdoor deployment should be even less dynamic• fewer path possibilities

Latest work: Silence with responsiveness•Organizationally informed routing

•How do agents need to communicate?•How often?•How much bandwidth?•How much responsiveness? •With what priority?

•Use organization knowledge to improve predictive routing

•Developed eCQRouting algorithm•Provides as much as 43% increase in application-level bandwidth over OLSR

•Details in our AAMAS-08 paper (presenting on Wed)

Latest work: Silence with responsiveness•Rapid radio cycling

•shorter-duration communication windows•much higher communication-cycle frequency•increased responsiveness with the same power expenditure

•each agent turns on radio•listens to see if it is needed for communication•if not, turn off radio (its window is over)

•“waves” of communication across network•messages don’t have to travel their full path in one cycle

•held to be forwarded in next cycle (occurring momentarily)

•challenge is maintaining routing with rapid cycling•how much “on-time” is needed to support path exploration?

•this work is just beginning…

Next-generation hardware

•PASTA is showing its age•new “ziti” processors are no longeravailable

•firmware issues remain unresolved

•Netgear MA111 (IEEE 802.11b) is outdated•Following Talisman-Saber we began looking into new hardware possibilities for CNAS agents•Connex & Verdex motherboards (Gumstix)

•commodity devices at commodity prices•not targeted for low-energy applications

•acceptable trade-off, as WiFi & GPS are the big spenders

Gumstix PXA processors

•Gumstix connex waysmall 400xm• Processor: 400MHz Intel XScale PXA255• Memory: 64MB RAM, 16MB Flash

•Gumstix verdex XL6P• Processor: 600MHz Marvell® PXA270 with XScale• Memory:128MB RAM, 32MB Flash

•Lower-cost hardware will enable permanent deployment of CNAS agents• cost of losing a sensor agent due to damage or theft becomes tolerable

CLISP on the Gumstix

•Unable to use the Debian ARM 2.34 package•software tightly coupled with kernel & buildroot version (as flashed on the Gumstix)

•Cross-compiled CLISP using linux/86•semi-automated process•former cross-compilation support in CLISP’s build scripts had fallen into disrepair

•Latest news:•support for fully-automated cross-compilation will be included in the next CLISP (2.45) release!•…sometimes a bit of positive complaining works!!

Summary

•CNAS has been successful •as a deployed agent-based sensor network•as a research environment

•agents operating near the limit of radio range with their radios turned off most of the time present different challenges than traditional MAS settings •only starting to incorporate more complex activities and adaptive reasoning into CNAS

•Working on CNAS has also been a lot of fun!

•we are already looking forward to Talisman-Saber 2009!

Questions & Discussion