INNOVATIVE MODIFICATION IN REINFORCEMENT LEARINING MODELS USING RECURNT NEURLA NETWORK AND AUTOECNODER

Abstract

The researchers explored building a generative model with the combination of Variational Autoencoders (VAEs) and Generative Adversarial Network (GAN) to offer better results when the agent interacts with the environment. The premise is that the agent model can train the unsupervised environment and increase the imaging quality. An experiment was performed on car racing based on the designed agent model with features that could effectively be extracted to help produce reliable information. The combination of VAE and GAN provides better feature extraction to gather relevant data and solve the complexity. Another novel approach was performed in which the world model has a modified Recurrent Neural Network (RNN) model and was carried out by a bi-directional gated recurrent unit (BGRU) as opposed to a traditional long short-term memory (LSTM) model. BGRU tends to use less memory while executing and training faster than an LSTM, as it uses fewer training parameters. However, the LSTM model provided greater accuracy with datasets using longer sequences. Based on practical implementation, the BGRU model produced better performance results.