Quest¶
Overview¶
The Quest algorithm has been ported directly from the MATLAB sources available with the PsychToolbox. The MATLAB source code was written by Denis Pelli. Thanks to Denis for allowing it to be released under the BSD license to the (Python) world.
This Python version does not depend on the Vision Egg, and may be useful in other contexts.
Download¶
Download the source directly from the Quest package at the SourceForge download page.
Example¶
The intensity scale is abstract, but usually we think of it as representing log contrast.
Specify true threshold of simulated observer: 0.5
Estimate threshold: 0.4
Trial 1 at 0.7 is right
Trial 2 at 0.4 is right
Trial 3 at -0.2 is wrong
Trial 4 at 0.7 is right
Trial 5 at 0.4 is right
Trial 6 at 0.4 is wrong
Trial 7 at 1.1 is right
Trial 8 at 0.9 is right
Trial 9 at 0.7 is right
Trial 10 at 0.7 is right
36 ms/trial
Mean threshold estimate is 0.60 +/- 0.38
Quest beta analysis. Beta controls the steepness of the Weibull function.
Now re-analyzing with beta as a free parameter. . . .
logC sd beta sd gamma
0.62 0.39 3.7 4.5 0.500
Actual parameters of simulated observer:
logC beta gamma
0.50 3.5 0.50
The example above is taken directly from the demo() function of Quest.py and is a direct translation of the demo included in the original MATLAB source:
print 'The intensity scale is abstract, but usually we think of it as representing log contrast.'
tActual = None
while tActual is None:
sys.stdout.write('Specify true threshold of simulated observer: ')
input = raw_input()
try:
tActual = float(input)
except:
pass
tGuess = None
while tGuess is None:
sys.stdout.write('Estimate threshold: ')
input = raw_input()
try:
tGuess = float(input)
except:
pass
tGuessSd = 2.0 # sd of Gaussian before clipping to specified range
pThreshold = 0.82
beta = 3.5
delta = 0.01
gamma = 0.5
q=QuestObject(tGuess,tGuessSd,pThreshold,beta,delta,gamma)
# Simulate a series of trials.
trialsDesired=100
wrongRight = 'wrong', 'right'
timeZero=time.time()
for k in range(trialsDesired):
# Get recommended level. Choose your favorite algorithm.
tTest=q.quantile()
#tTest=q.mean()
#tTest=q.mode()
tTest=tTest+random.choice([-0.1,0,0.1])
# Simulate a trial
timeSplit=time.time(); # omit simulation and printing from reported time/trial.
response=q.simulate(tTest,tActual)
print 'Trial %3d at %4.1f is %s'%(k+1,tTest,wrongRight[response])
timeZero=timeZero+time.time()-timeSplit;
# Update the pdf
q.update(tTest,response);
# Print results of timing.
print '%.0f ms/trial'%(1000*(time.time()-timeZero)/trialsDesired)
# Get final estimate.
t=q.mean()
sd=q.sd()
print 'Mean threshold estimate is %4.2f +/- %.2f'%(t,sd)
#t=QuestMode(q);
#print 'Mode threshold estimate is %4.2f'%t
print '\nQuest beta analysis. Beta controls the steepness of the Weibull function.\n'
q.beta_analysis()
print 'Actual parameters of simulated observer:'
print 'logC beta gamma'
print '%5.2f %4.1f %5.2f'%(tActual,q.beta,q.gamma)
References¶
- Watson, A. B. and Pelli, D. G. (1983) QUEST: a Bayesian adaptive psychometric method. Percept Psychophys, 33 (2), 113-20.
- Pelli, D. G. (1987) The ideal psychometric procedure. Investigative Ophthalmology & Visual Science, 28 (Suppl), 366.
- King-Smith, P. E., Grigsby, S. S., Vingrys, A. J., Benes, S. C., and Supowit, A. (1994) Efficient and unbiased modifications of the QUEST threshold method: theory, simulations, experimental evaluation and practical implementation. Vision Res, 34 (7), 885-912.