All numeric parameters for a map are computed from the raw segment data
of other maps using leave-one-map-out cross-validation. First the actual
values of the relevant properties (e.g., length, depth, angle) are computed,
and then parameter values are computed by a function of the mean and standard
deviation of these values.
//////////////numeric parameters definitions begin//////////////////
DoorLength = the mean of the length of all Door segments.
DoorDepth = the mean of the depth of all Door segments.
WallLength = the mean of the length of all Wall segments.
WallDepth = the mean of the depth of all Wall segments.
MaxDoorDepth = mu + 1.5*sigma, where miu is the mean of
the depth of all Door segments, and sigma is the standard deviation
MinWallDepth = mu - 1.5*sigma, where miu is the mean of the depth, (which is
typically around 0) of all Wall segments, and sigma is the standard deviation
MinDoorLength = mu - 1.5*sigma, where miu is the mean of the length
of all Door segments, and sigma is the standard deviation
MaxDoorLength = mu + 1.5*sigma, where miu and sigma is similar as above
MinWallLength is the lower bound of the length of Wall segments
MaxAngle is computed by mu + 1.5*sigma, where miu is the mean of the
angle values of all Wall segments, and sigma is the standard deviation
//////////numeric parameters definitions end///////////
////////parameter settings for learning (VP + HMWS) begin/////////
number of learning iterations: 100
global learning rate: 1.0
//HMWS settings are same as HMWS inference setting
//////////parameter settings for learning (VP + HMWS) end/////////
//////////parameter settings for HMCS begin//////////////////////
num of HWS steps for each HSS step: 100000
probability of random move for HWS: 0.5
probability of simulated annealing for HSS: 0.4
simulated annealing temperature: 0.5
number of steps of HSS after reaching a solution: 100
number of burn-in steps: 1000
///////////parameter settings for HMCS end////////////////////////
/////////////parameter settings for MCS begin/////////////////////
num of WS steps for each SS step: 100000
probability of random move for WS: 0.5
probability of simulated annealing for SS: 0.5
simulated annealing temperature: 0.5
number of steps of SS after reaching a solution: 100
number of burn-in steps: 1000
/////////////parameter settings for MCS end///////////////////////
//////////parameter settings for SMTP begin////////////////
SMTP number of runs: 3
SMTP number of swapping chains: 10
SMTP chain weight: log distributed, i.e., w, w/k, w/k^2, ..., w/(k^n-1).
k is the (n-1)th root of the largest weight
initilization for each chain: by MWS for hard clauses
number of burn in steps: 1000
////////////parameter settings for SMTP end///////////////
/////////////parameter settings for Gibbs begin///////////
Gibbs number of chains: 10
initilization for each chain: by MWS for hard clauses
Gibbs samples per test: 100 (actually it doesn't matter
because the inference never converges during experiments)
number of burn-in steps: 1000
/////////////parameter settings for Gibbs end////////////
///////////parameter settings for HMWS begin////////////
probability of random move: 0.06
max number of flips: 100000
max number of retries: 10
standard deviation for random move of numeric variables: depends on data
///////////parameter settings for HMWS end/////////////
///////////parameter settings for MWS begin////////////
probability of random move: 0.5
max number of flips: 100000
max number of retries: 10
///////////parameter settings for MWS end/////////////
////////////parameter settings for SA begin///////////
initial temperature: 10
temperature reduction rate: 0.96
number of steps for every temperature reduction: 100
standard deviation of numeric variables for SA proposal distribution:
depends on data
////////////parameter settings for SA end////////////