1. I am measuring the resistance to 0.1 intervals. (from 350 ohm to 700ohm)
Like I am reading 350.1 , 500.6 etc.

But I have to deliver accuracy to two decimal places.


2. Is it ok , if I take 256 samples of reading with 0.1 interval, & average them to get second decimal place also.

 

You can (sometimes) increase resolution and the accuracy of the conversion process by oversampling.

http://en.wikipedia.org/wiki/Oversampling#Resolution
http://www.microchip.com/stellent/idcplg?IdcService=SS_GET_PAGE&nodeId=1824&appnote=en533730

 

No one can tell what the accuracy of the measurement is from that tiny bit of information.

What is the circuit?
Is there a stable reference voltage?
Good layout with little noise?
What A2D converter, and how many bits?

Oversampling can sometimes get more digits, but there must be noise in the measurement and it has to be white Gaussian noise (or something like that, don’t quite remember as I don’t ever do this). If you simply sum 256 of the same reading you have not added any accuracy.

Without seeing your circuit I would be surprised if your resolution (which is not accuracy) is less than 1 ohm anyway.

 

Accuracy is dependent on the reference voltage used for the ADC and the number of bits of the ADC. For example, a ten bit ADC will give 1024 steps of the reference voltage. If the reference voltage is 5V, the minimum step is about 0.005V

What are you using to display your output? It could also be a programming limitation.

 

Accuracy is dependent on the reference voltage used for the ADC and the number of bits of the ADC. For example, a ten bit ADC will give 1024 steps of the reference voltage. If the reference voltage is 5V, the minimum step is about 0.005V...

Oversampling, as suggested above, can increase the precision (resolution) and DC accuracy.

But as Ernie said, uncorrelated noise must be present (or injected) into the signal.

Ideally, the noise should contain all frequency components up to the nyquist criterion (1/2 the sample rate), with a peak-to-peak value of 1/2 lsb, IIRC.

 

to read to 1/10 ohm, you will have to take extreme care with your connections and lead resistances too. it dosnt take much test fixture (probes and ect.) to make changes in readings.

 

Oversampling, as suggested above, can increase the precision (resolution) and DC accuracy.

But as Ernie said, uncorrelated noise must be present (or injected) into the signal.

Ideally, the noise should contain all frequency components up to the nyquist criterion (1/2 the sample rate), with a peak-to-peak value of 1/2 lsb, IIRC.

When I was goofing around with detecting moon power on a solar panel I used a few methods to sample a 12-bit ADC to generate the graphing data.
The first night was just an average of 128 samples and a smoothing filter.
The second night was with 128 samples and 2 bit resolution extraction using an simple digital oversample filter for 14-bit effective resolution and increased accuracy of the signal using the noise signal. (The next night clouds rolled in at about 200 data point so you can see the small changes in the thin cloud layer here).

With a fairly (random thermal type) noisy signal it works but I don't think it's what the OP really needs as others have suggested.

 

Accuracy and resolution are two different things. Resolution is how many digits you get. Accuracy determines how many of these digits are accurate. In most real life applications, resolution is far greater than accuracy. Therefore, it is very unlikely that by increasing resoltion you get any gain in accuracy.

In your example, you get 0.02% resolution, and you want to increase it to 0.002% resolution. Unless you build very accurate (and expensive) circuit, your accuracy is unlikely to be better than 0.1% (more likely around 1% or so).

Oversampling increases resoltion (proportional to N). It also may increase accuracy by reducing random noise (proportional to sqrt(N)). But it is absolutely helpless against any kind of systematic errors (thar's what most of the real errors are).