$ docker run -ti poker-bot:1.0 -env leduc Playing in the Limit Hold'em environment: $ docker run -ti IMAGE_NAME:TAG -env leduc Playing in the Leduc Hold'em environment:

After building the image, which gives a simple containerization of our application, the game runs successfully in its container.įor building the image use following command:Į.g. The Dockerfile contains the list of system dependencies. You can choose between running code using Docker or using Colab notebooks. The playing game code runs in the Leduc Hold'em environment by default. These pre-trained agents can be set as opponents in the Leduc and Limit Hold'em environment. In the final code, we saved the best agents after hyperparameter optimization. These references were used during the implementation of the DQN agent in PyTorch. These can be set in the training code for the DQN agent. The agent can be trained and evaluated against a random agent and a pre-trained agent. If action Fold has the maximum value, we use Raise action if possible If action Check has the maximum value, we use Raise action if possible If action Call has the maximum value, we use Raise action if possible Using action with maximum value (default in DQN) The possible settings are displayed below: In case of the given action has the maximum q-value, the agent chooses the Raise action instead if it is a valid action. These features are described in the Nature paper Human-level control through deep reinforcement learning.įurthermore, as an extra component, we added the opportunity of a more aggressive playing strategy. Second, to make the training more stable, another Q-network is used as a target network in order to backpropagate through it and train the policy Q-network. First, it uses a replay buffer to store past experiences and we can sample training data from it periodically. This implementation is an advanced Q-learning agent in two aspects. We used the RLcard DQN agent written in TensorFlow as a base and created a more powerful, more manageable, and easy to use code in Pytorch. The code for the second milestone is a DQN agent in PyTorch. It is used as a presentation that the chosen environment works and the agent is ready to train. The presented code for the first milestone is based on the RLcard github repository example code. Team members: László Barak, Mónika Farsang, Ádám Szukics After training, we can play against our pre-trained agent. Our project focuses on reinforcement learning with the aim of training an agent in a poker environment. This repository contains the project for the Deep learning class (course code: VITMAV45) at the Budapest University of Technology and Economics.

Code the wheel poker bot