Search All of the Math Forum:
Views expressed in these public forums are not endorsed by
Drexel University or The Math Forum.



Re: neural network
Posted:
Jan 19, 2013 9:34 PM


"Jamaa Ambarak" <jamaa73@yahoo.com> wrote in message <kd49q6$c8u$1@newscl01ah.mathworks.com>... SNIP > > Dear Greg, > > the training data for example > > let we say that there are 15 drivers here and every driver has his own number of lane deviation and they vary e.g d1 has 50 times , d2 has 300, d3 has 400 times ,,,,,,,d15 has 200 times out of lane. the drift is very also some times one second or 2 second and after that the driver take action. so the most important for me the first sample of the lane deviation this happen at lateral position = /+ 0.81 m and then I cut out the rest. so I have two kinds of windows ''1'' or ''0'' , the first '1' window start one second = 50 samples before the lane happen, and the '0' start after the driver go back to the lane and its same size of window one.the 1 window is [150X50], inputs are 50 samples lateral position, 50 samples speed velocity, and 50 samples steerangle. and the target is 1 it means out of the lane, as well the 0 window but has target =0. then the total 1 windows and 0 windows > > for driver d1 100 windows 50/50 in and out. so as well for the other drivers. the > >training data I did select 10 drivers for training and 5 drivers for testing.
Less bias if you average over 3 trials so that each driver is used twice for training and once for testing.
> > for testing data example d1 has 50 times lane deviation > the total length of lateral position is 400320 after I cut the duration when the drivers >out of the lane as well same length for other features...
Training data:
The 150dim input vector corresponding to a "1" target is one second of 3dim data (@ 50 samples/sec) with lateral deviation < 0.81m JUST BEFORE a continuous interval >= 1sec with lateral deviation >= 0.81 m.
Correct?
The 150dim input vector corresponding to a "0" target is one second of 3dim data (@ 50 samples/sec) with lateral deviations < 0.81 m JUST AFTER a continuous interval >= 1sec with lateral deviations > 0.81 m.
Correct?
Testing
The 150dim input vector corresponding to a test target is a one second sliding window of 3dim data (@ 50 samples/sec).
Correct? > > here is the code > > %% load inputs and targats data > > Ptemp = P; random 1's or 0's windows > > Ttemp = T; (0 or 1) > > rnd=randperm(3858); > > for ii=1:length(rnd) > > P(:,ii)=Ptemp(:,rnd(ii)); > > T(:,ii)=Ttemp(:,rnd(ii)); > > end
If they are originally random, why do you have to randomize them again?
> > load y21; testing files > > A=y21; > > % mapping and normalizing
You might want to try it without mapping and normalizing
> > [pn,ps]=mapminmax(P); > > [tn,ts]=mapminmax(T); should I normalizing the target or no
I see no reason to normalize the targets. {0,1} is fine and should be used with a logsig output transfer function.
> > [an]= mapminmax('apply',A,ps); > > % %% create network (3 layers with 6 nodes 10 nodes)''
There is no reason to use more than one hidden layer
> >traingdm '','trainscg','trainlm',trainbfg) ,,,, which one should I use as traing function
Use 'trainscg' for classification problems with {0,1} targets. Otherwise use the default 'trainlm' unless there are speed/memory problems.
> > % net = newff(minmax(pn),[8 6 1],{'tansig','tansig','purelin'},'traingdm'); > > which one should I use as activation function..
rand(0) % Initialize the RNG so you can duplicate runs net = newff(minmax(pn),[ H 1],{'tansig','logsig'},'trainscg');
Optimize a minimum successful H using trial and error.
> > % % view(net)
delete below to use as many defaults as possible
> > % net.trainParam.epochs=1000; > > % net.performFcn = 'mse'; > > % neuronal_net.trainParam.show =50; > > % net.trainParam.goal = 1e6; > > % net.trainParam.lr=0.3; > > % net.trainParam.min_grad =1e6; > > % net.efficiency.memoryReduction; > > % net.trainParam.mc=0.6; > > % net = init(net); > > % % net=train (net,P,T);
delete above to use as many defaults as possible
net.trainParam.goal = 0.01*mean(var(T)) net.trainParam.show = 10
loop over candidate values for H (10 trials each) to optimize H and random initial weights
> > % [net,tr] = train(net,pn,tn);
[ net, tr, Y, E ] = train(net,pn,T); % Yn = sim(net,pn); E = TYn
I'll let you figure out the rest.
Greg
> > % t_pred=sim(net,pn); > > % y=sim(net,an); > > % y1 =mapminmax('reverse',y,ts); > > % % y1 =mapstd('reverse',y,ts); > > % predicted = hardlim(y1'  0.5); > > please read my notes beside the code and please tell me is it correct or no, I have training the network and give me %100 Sensitivity and %98 Specificity for most testing drivers



