Sponsored by: National Science Foundation #XX
Cognitive Engine for Adaptive Modulation
August 7th, 2009
REU program on Justin Waugh and Daniel González-Pérez
Create a cognitive engine which changes cluster size of a Quadrature
Amplitude Modulated (QAM) symbol in order to optimize for throughput and
reduce bit error rate based on signal to noise ratio (SNR).
•Starts off with a lookup table with SNR thresholds of when to change the
cluster size based on theoretical models to maximize for throughput (Figure 1)
•It has a goal Bit Error Rate to meet, and it changes the thresholds in order to
achieve this Fig. 1. Throughput of various constellation size signals Fig. 2. Comparison between simulated transmission and
in an AWGN channel. From top to bottom along right side theoretical model – Bold lines are theoretical bit error
•The theoretical models for the Bit Error Rate are shown on Figure 2 of graph they are: 512, 256, 128, 64, 32, 16, 8, 4 – QAM rates, while the thin lines are from the simulation. From
signals. left to right the constellation sizes are: 4, 8, 16, 32, 64, 128,
In the simulations we set up, we changed the SNR of the simulated channel in
three different ways and monitored the transmission packet-by-packet.
Equation used to model BER vs. SNR for M-QAM signal
Constellation Size Analyzed Signal
1 Random Fluctuating Noise (Figure 3)
•Constellation Size correlates with SNR.
•BER decreases down to the target (10^-3) even against the random noise.
Fig. 3. Plots of BER, SNR, Constellation Size, and Noise from the first simulation.
2 Step Function Noise (Figure 4)
•Periodicity in the Constellation Size.
•Engine tries to match goal BER, and if signal is too good it increases the
constellation size which causes error, subsequently making it decrease.
•The average BER for the entire transmission matches the goal BER.
Fig. 4. Plots of BER, SNR, Constellation Size, and Noise from the noise that acts like a step function.
3 Periodic Ramping Function Noise (Figure 5)
•First few periods the engine’s guesses are sub optimal resulting in lots of error.
(but it is still learning and approaching the goal)
•After the 3rd period the error in the signal has reduced to match the goal BER.
•By changing the modulation type at the right time the engine is able to attain
the requested goal BER.
Learned Responses (Figure 6)
The Engine learns new critical changing points for the modulation type
based on the environment it has observed. As shown below the learned response Fig. 5. Plots of BER, SNR, Constellation Size, and Noise from the simulation with periodic noise.
for different SNR simulations is very different.
•The cognitive engine was able to learn about its environment and by just
changing its cluster size achieve a specified bit error rate
•By changing the rating system used by the engine to change the thresholds and
controlling the power and frequency, its is theoretically possible to create the
best transmission through a channel, and adapt itself to any new environment
Daniel Amaury González-Pérez Justin Waugh
Fig. 6. Comparison between the engine’s learned switching points (bold blue line is cognitive engine’s results). University of Puerto Rico at Mayagüez Virginia Tech
Left, learned response for step function. Right, learned response for the periodic function
Email: firstname.lastname@example.org Email: email@example.com