Reinforcement Learning (RL) and in particular Deep Reinforcement Learning (DeepRL) are popular learning techniques. given the complexity and high dimentionality of such systems makes it difficult to provide any assurences about the system behavior. Moreover, given that most of the responsibility of the behavior is delegated to the neural network, which is a black box for programmers, evaluating and testing the algorithms is not transparent.
In this project we propose to build a testing framework for RL algorithms. The idea of the project is to propose a unifying theory to test different properties of RL algorithms (e.g., fairness, correctness, convergence, bias) for different DeepRL algorithms (e.g., DQN, DDQN, A3C, PPO, T3PO).
To generate the testing fraemwork for DeepRL we will take inspiration from existing testing approaches for RL or general Deep Neuran Network ML algorithms. based on these approaches, we will build a unifying theory to allow us to test different properties over a range of DeepRL algorithms.
Initially, we will evaluate the appropriateness of existing techniques on DeepRL algorithms. Then based on the imposibilities of the tecniques, we will propose extensions over the theory to cover as many algorithms possible, and as many system properties as possible.
n.cardozo