More talks in the program:
14:00 - 14:45
Google’s AlphaGo uses reinforcement learning to win against Go Masters. This method is especially useful when the search space of solutions is polynomial: at some stages in a game of Go, there are more possible moves than atoms in the universe. Instead of brute force, the algorithm uses reinforcement learning to do probabilistic search of potential solutions and make the move that’s most likely to result in a win. At Diffblue we use the same approach to write unit tests for Java projects, using reinforcement learning to search for tests that achieve coverage and usefulness goals, while remaining human-readable. The result is a suite of regression unit tests that run quickly and early to find bugs, accelerating DevOps pipelines and reducing cycle times. In this talk I’ll look at how machine learning works, supervised vs. unsupervised approaches, reinforcement learning and how it is applied in both in AlphaGo and Diffblue Cover.