Artificial intelligence has become the quickest developing field of computer science – on daily basis numerous publications report the latest results of various research groups. Applied artificial intelligence is becoming a part of our lives, as artificial decision-making is becoming more and more easily and widely applicable. Traditionally, learning algorithms require a data-driven approach, meaning that they are limited to those fields, where large amount of clean data (i.e. structured, with minor ambiguity) is available. Reinforcement learning yields different approach as it does not require expensively annotated, abundant data, since data is generated on the fly. This can be executed with an interactive agent recording its observations during wandering within an explorable environment. These pieces of recordings become the training data later on.
In my thesis, I implement reinforcement learning using artificial environments, which are computer games in my case. In reinforcement learning game-winning strategies can be explored via applying arbitrary actions. Based on the explorations, this random policy can be iteratively adjusted, resulting in intelligent behavior. My goal is to implement an agent that can learn the nature of a particular game through its experiments, and as a result be able to become an intelligent player.
The implemented solution can be divided into three milestones. After the constructing the network of relationships between different frameworks and libraries, I have implemented deep learning based intelligent agents, and at last I have trained those agents to be able to beat the games. The training has been executed on two different environments (an open and a closed one).