Project

Implementation of a neural network agent that plays video games using reinforcement learning

This project to implements a generalized neural network agent that plays different
 video games using reinforcement learning algorithm. This project uses OpenAIs
 simulated video game environment ‘gym’ for training and testing the proposed
 reinforcement learning algorithm solution. For training the neural network agent,
 different neural network models are used like Convolutional Neural Networks, Recurrent
 Neural Networks and combination of both. Python and TFlearn (Tensorflow backend)
 are used to implement the project. The results show that the proposed solution works well
 for an average of two to three games. However, performance of the solution is degraded
 when neural network is trained on four or more video games. Although using the ‘TopK’ metric (which is added to the proposed solution to increase the efficiency of neural
 network to play multiple video games) yields a dramatic increase in training and
 validation accuracy of the neural networks, the networks are still not able to play variety
 of video games with good degree of precision. To improve the performance, deeper neural
 network models like VGG-19 can be utilized in the future, given that hardware resources 
 required by such models are available (e.g., a GPU with a larger global memory is
 needed).

Project (M.S., Computer Science)--California State University, Sacramento, 2018.

This project to implements a generalized neural network agent that plays different video games using reinforcement learning algorithm. This project uses OpenAIs simulated video game environment ‘gym’ for training and testing the proposed reinforcement learning algorithm solution. For training the neural network agent, different neural network models are used like Convolutional Neural Networks, Recurrent Neural Networks and combination of both. Python and TFlearn (Tensorflow backend) are used to implement the project. The results show that the proposed solution works well for an average of two to three games. However, performance of the solution is degraded when neural network is trained on four or more video games. Although using the ‘TopK’ metric (which is added to the proposed solution to increase the efficiency of neural network to play multiple video games) yields a dramatic increase in training and validation accuracy of the neural networks, the networks are still not able to play variety of video games with good degree of precision. To improve the performance, deeper neural network models like VGG-19 can be utilized in the future, given that hardware resources required by such models are available (e.g., a GPU with a larger global memory is needed).

Relationships

Items