Algorithmic trading strategies are automated defining a sequence of instructions executed by a computer. A good strategy should be profitable which includes identification of what to trade and how to trade. In this paper, we focus on the study of algorithmic trading strategy optimization and propose a strategy optimization model based on an initialized strategy pool. In order to get a better strategy, a mutual information entropy based clustering algorithm is employed to analyze the correlations among the stocks and a reward and punishment scheme is also set up for updating the latest transaction data in the strategy optimization process. Experimental results on several different groups of stocks showed that in most cases, this optimization model can find a profitable strategy swiftly.