The exploration exploitation dilemma is one of the most prominent open research problems in Reinforce�ment Learning (RL). Pathak et al [67] have proposed a solution that introduces balance between those two components. The solution involves an agent, that is the RL algorithm , producing its own intrinsic rewards, through a bolt-on module, namely Intrinsic Curiosity Module(ICM). Using intrinsic rewards the algorithm is guided towards more exploratory behaviour but still not loosing sight of its ultimate goal of maximising its learning progress, which is quantified by the extrinsic rewards it has accumulated by the end of the training. The first aim of this project is to implement this solution in environments with dense discrete extrinsic rewards supplied to the agent and importantly modify the implementation so that it works in Atari environments. The results of the implementation align with the one;s of the original creator of the ICM, as it is deduced that intrinsic curiosity boosts the learning progress of the agent. Moreover, as it has been suggested by prior research a form of intrinsic reward is being produced in the brain, which is in this research hypothesis it is proposed that it is processed in the same way as the ICM. The ICM is then paralleled with the functionality of the neocortex which is believed to be the main component of the brain that drives learning and is responsible for functionalities that distinguish humans from other mammals. The results of this implementation are in a promising direction as it is illustrated that the more biologically plausible model does yield even better results than the original ICM. However, this must be subject of further research as the reasons underlying the boost in learning performance of the more biologically plausible ICM compared to the original ICM remain in debate
• Python as a programming language to implement the ICM.
• PyTorch Python’s Machine Learning Framework to implement the ICM, this include using convolutional neural networks for encoding features of incoming images and neural networks for making predictions.
• OpenCV computer vision library to process incoming raw input images so that computations are more efficient.
• Open-AI Gym open source toolkit containing a range of environments where the agent can learn, but also allowing to directly modify the open-AI Gym environment by using Gym wrappers.
• Using Google's DeepMind A3C agent as a baseline