engleski

Reinforcement learning in simulated systems

In this thesis we focused on implementing a reinforcement agent that was able to generalize to various simulated systems and in doing so ; showed the ability of the reinforcement learning algorithm to adapt. Fundamentals of deep learning were also described. Using Tensorflow, a reinforcement learning agent was implemented. The agent was evaluated on three different scenarios: Pacman from the Atari 2600 games and two different scenarios of the game Doom. The evaluation results were shown via graphs.

Reinforcement learning ; agent ; environment ; deep neural networks ; Q values ; state values ; experience replay ; recurrent networks.

nije evidentirano