Title: Wireless Link Adaptation for Deadline Constrained Traffic with Imperfect Channel Estimates
1Wireless Link Adaptation for Deadline Constrained
Traffic with Imperfect Channel Estimates
Tim Holliday Depts of MSE and E.E. Stanford
University thollida_at_stanford.edu
Andrea Goldsmith Dept of E.E. Stanford
University andrea_at_ee.stanford.edu
Peter Glynn Dept of MSE Stanford
University glynn_at_stanford.edu
IEEE ICC 2002 New York, NY
2The Problem
- Streaming audio/video and other deadline
constrained traffic require tight QoS constraints - Current Fixed-BER control policies consume far
more power than necessary and reduce system
capacity - Delayed and erroneous channel estimates can wreak
havoc on the performance of Fixed-BER controls - High speed mobiles and urban propagation
environments are a particularly bad combination
3The Main Issues
- How do you solve for optimal link adaptation
policies for deadline constrained traffic? - Minimum power consumption subject to constraints
on deadline expiration - How do channel characteristics and poor channel
estimates affect the optimal control? - What types of QoS constraints accurately
represent performance? - Is a simple constraint on probability of data
loss sufficient? (Answer It depends on the
channel)
4A Solution
- We propose a dynamic programming solution for
optimal power and coding with deadline
constrained traffic
- Key advantages
- Permits a great deal of flexibility and detail in
our wireless channel models - Allows for a wide variety of performance
constraints - Allows us to adapt to delayed/inaccurate channel
estimates
5System Model
- Standard TDMA cellular structure (no intra-cell
interference) - Single wireless mobile transmitting deadline
sensitive data to a base station - All data generated at the mobile has a deadline
by which is must be transmitted - Mobile can adapt transmission power, modulation,
and convolutional code parameters to meet
deadline requirements
6Dynamic Programming Formulation
- Construct a Markov chain model for a mobile
- See the paper for details (Markov modulated
traffic sources, traffic with multiple deadlines,
etc.) - Each unique combination of power, modulation, and
coding determines a different transition matrix - Objective of the DP is to choose the optimal
transition matrix for each state of the mobile
7Value Function and Transition Dynamics
- Define a control policy g as a function that maps
states into actions (i.e. a transition matrix) - Then define a value function and transition
matrix (V(g) and P(g)) such that
8Optimization and Constraints
- We can optimize the value function through a
simple linear program. - This structure also allows us to easily constrain
various metrics like probability of deadline
expiration
9Imperfect Channel Estimates
- Consider a simple two-state Markov chain channel
model with transition matrix Q. - Suppose the estimates are uncertain and delayed
by 5 timeslots. The best guess of the channel
state is NOT the old estimate!
10Numerical Example Setup
- System is modeled after EDGE, see the paper for
details on time slot structure, BLER plots, etc. - Mobile adaptation options
- Transmission power range 20mW to 800mW
- Choose between MCS-6 and MCS-9 (rate ½ and full
rate 8PSK) - Traffic deadlines are 60ms or 3 time slots
- Deadline constraint is 1
11Channel Models
- Consider two mobile speeds 3 km/h and 50 km/h
- Two shadowing environments with mean path loss of
120dB - Macrocell
- Std. Dev. 7dB
- Correlation.82 at 100 meters
- Microcell
- Std. Dev. 4.3dB
- Correlation.3 at 20 meters
- So we should expect estimate delay to have bigger
impact in a 50 km/h microcell than a 3 km/h
macrocell
12Power Cost For A Fixed-BER Control
13Power Cost For Dynamic Programming Solution
14Example Control Policy
153 km/h
163 km/h
50 km/h
17Power Cost With Consecutive Loss Constraint
18Conclusion
- We have presented a framework for finding optimal
link adaptation policies for deadline constrained
traffic - We also proposed a means to adapt to channel
estimation errors and delays - Both the optimal control and the validity of the
QoS constraints can be greatly affected by the
channel model