Hough transform has been the most common method for circle detection exhibiting robustness but adversely demanding a considerable computational load and large storage. Alternative approaches include heuristic methods that employ iterative optimisation procedures for detecting multiple circles under the inconvenience that only one circle can be marked at each optimisation cycle demanding a longer execution time. In contrast, learning automata (LA) is a heuristic method to solve complex multi-modal optimisation problems. Although LA converges to just one global minimum, the final probability distribution holds valuable information regarding other local minima which have emerged during the optimisation process. The detection process is considered as a multi-modal optimisation problem, allowing the detection of multiple circular shapes through only one optimisation procedure. The algorithm uses a combination of three edge points as parameters to determine circles candidates. A reinforcement signal determines whether such circle candidates are actually present at the image. Guided by the values of such reinforcement signal, the set of encoded candidate circles are evolved using the LA so that they can fit into actual circular shapes over the edge-only map of the image. The overall approach is a fast multiple-circle detector despite facing complicated conditions.