The master’s thesis deals with the development and implementation of an environment for deep reinforcement learning of market-making strategies. The theoretical foundations of Markov decision processes and deep reinforcement learning techniques are presented, with emphasis on $Q$-learning and deep $Q$-learning. The work includes a detailed analysis of the limit order book and various types of market orders, which are key to understanding the dynamics of financial markets. We created a simulation environment for testing trading strategies, where we first implemented an agent on a simplified market using tabular $Q$-learning and deep $Q$-learning. The analysis of results showed the effectiveness of different approaches depending on environmental parameters. The developed agent was then adapted for trading on real crypto market data using technical indicators and more advanced neural network architectures. Experimental results show that strategies learned through deep reinforcement learning can successfully compete with traditional market-making approaches in various market conditions.
|
