A simulator for evaluating machine-learning algorithms for autonomous ships

In this thesis, a ship simulator capable of training agents in ship behavior is presented. The simulator can be used to train and simulate agents with different kinds of behaviors. The simulator supports ships and boats of different sizes. Users can input configuration files with agent and environment data to create their own scenarios. The agent data includes ship information such as the position, speed and heading of the vessel as well as the weights of the training data which define the behavior of the agent. The environment data allows the user add a map or obstacles as well as decide the size of the area the agents move on. The simulator is capable of training and simulating multiple agents simultaneously. The simulators time is divided into one second long time steps. The application does not work in parallel but instead handles all calculations between each time step. With various configuration options the simulator can be used to simulate a plethora of scenarios. New reinforcement learning methods and ship configurations can also easily be added. The simulator can be used in educational and industrial applications in reinforcement learning or autonomous surface vehicle navigation. The simulation is shown by a simple top down view. All the objects are two-dimensional polygons and therefore, the ships motions are composed of 3 degreesof-freedom: surge, yaw and sway