Learning automaton （LA） is an adaptive decision maker that learns to choose the optimal action from a set of allowable actions through repeated interactions with a random environment. In most of the traditional LA, the action set is always taken to be finite. Hence, for continuous parameter learning problems, the action space needs to be discretized, and the accuracy of the solutions depends on the level of the discretization. A new continuous action set learning automaton （CALA）is proposed. The action set of the automaton is a variable interval, and actions are selected according to a uniform distribution over this interval. The end points of the interval are updated using the binary feedback signal from the environment. Simulation results with a multi-modal learning problem experiment demonstrate the superiority of the new algorithm over three existing CALA algorithms.