You don't need history prices, but you need a discounting curve.

1) Simulate the price of S as what you do for European option (i.e., dS = rSdt + sigma*S*dX).
2) Using LONGSTAFF & SCHWARTZ REGRESSION to make adjustment for American options.

See Paul Wilmott on Quantitative Finance Ch 80 for details.

Personally I don't recommend to use MC for American option, esp when the dimension is low.
Maybe consider to use FD or tree instead?

By the way, if it's a call option on a stock with no dividend-paying, it won't early exercise.

Thanks a lot.
you do make sense about my question,however, what you refers is just model-based (such as BS-model) ,so you can use LS-MonteCarlo to generate the paths, not dependent on historical prices.
but, i am pricing american options  is model-free, only using the market prices. the paths simulation is required when using Longstaff method(2001),and this is the point i am encountering.
how can i get the simulation paths?

You are right.
My proposal is based on BS model.

I am thinking there is no pricing method which is model-free, as model is always needed to simplify future world (i.e., price movement in this case). Using historical prices to project future prices is a model assuming that the future will follow the historical pattern.
One thing you might notice is that arbitrage will exist too if you use historical prices to calculate option value.

Please freely make comment. :)

1. for BSmodel, it's easy to simulate many  prices paths, that's,we can  take random number.
2. but now, I have only 1000 historical prices, assuming the future prices follows the historical pattern, how can i produce 10000 paths? just using those 1000 data via the order "randsrc" in matlab?