An Empirical Assessment of the Performance of Multi-Armed Bandits and Contextual Multi-Armed Bandits in Handling Cold-Start Bugs

Bug triaging is a crucial activity in software development that involves identifying and prioritizing bugs for fixing. But recommending cold bugs to an appropriate developer is a more challenging problem. In this work, an empirical assessment of the effectiveness of two reinforcement learning algorithms, Multiarmed bandits and Contextual bandits, in the context of triaging cold-start bugs, is investigated. Five publicly available open-source bug-triaging datasets have been used to evaluate the performance of the algorithms using two standard evaluation metrics i.e., rewards and average rewards. Evaluations are done of two MAB algorithms- ɛ-Greedy and UCB, and LinUCB as a Contextual MAB algorithm. Our results showed that both Multiarmed bandits and Contextual bandits are effective in triaging cold start bugs but in different settings. Contextual bandits outperformed Multiarmed bandits in terms of rewards and average reward in different simulation settings. Our results demonstrate that contextual MAB algorithms outperform traditional MAB algorithms in accurately recommending cold bugs to an appropriate developer.