Test Case Prioritization (TCP) based on the continuous decision of Reinforcement Learning (RL) has achieved a successful application for test cases optimization in Continuous Integration (CI). The reward functions of RL describe how a test case "ought" to be executed in next integration, of which the design is usually based on the historical executions of the test case. The Average Percentage of Historical Failure (APHF) had been considered as one of the best reward function which has a strong correlation with the recent failure executions of a test case. However, for a test case with many historical failures but passes in recent cycles, the APHF value may be low. In this paper, two novel reward functions are proposed focusing on the impact of failure position in test case history execution sequence, which are the Average Position Exponential Weight (APEW) reward function and the Average Position Quadratic Weight (APQW) reward function, respectively. Both APEW and APQW carry out weight design of failure position but with different weights. We theoretically prove the issue of the only strong correlation with recent failure executions, and also prove that both proposed reward functions can reflect the quantity of historical failures and the distribution of these failures. Experimental verification on 10 industrial-level data sets show that the proposed reward functions can effectively improve the fault detection capability of test cases.

In the continuous integration, test case prioritization can effectively alleviate the resource-intensive problems associated with frequent integration commits. Test case prioritization in continuous integration is a sequential decision problem from which reinforcement learning is applied and can effectively adapt and learn from a changing environment. However, continuous integration testing brings new problems of sparse rewards to reinforcement learning because of frequent integration with low test failure and this problem can be addressed by increasing the number of rewarded test cases. In this paper, we propose a reinforcement learning reward object selection strategy based on Test Case Synchronization and Diversity (TCSD) that rewards failed test cases and with an additional selection of passed test cases with potential failure ability. The experiments on six real-world industrial data sets show that TCSD improves the learning efficiency and fault detection ability of reinforcement learning 6.35% in average NAPFD compared with the traditional strategies.