Title: Towards Automated Decision-Making Through Reinforcement Learning
Abstract:
Artificial intelligence (AI) holds considerable promise to tackle challenging decision-making problems such as automated health testing and drug discovery. Among the different AI paradigms, reinforcement learning (RL), a technique that concerns autonomous agents interacting with their environment in the hope to maximise the expectation of a reward signal, has achieved considerable milestones in real-world problems, leading to multiple state-of-the-art solutions. Yet, modelling problems in a fashion where RL is applicable may be non-trivial and require of careful thinking, e.g., if we are building an agent that learns to detect Alzheimer from brain scans, how do we design the problem so that this agent can “take actions” on the given images? and how do we reward such actions so that the agent learns to tell whether there is Alzheimer? Moreover, different families of RL algorithms hold diverse strengths and biases, meaning that evaluating which kind of algorithms is best suited for our problem may be key to success. We will go through all these aspects within RL and continue by providing evidence of how brittle RL solutions can be if we do not take the necessary steps towards robustness and generalisation. Last, for those interested in basic research, we will see some comparisons between RL agents and animal cognition and will review basic tools to test RL solutions before bringing them to costly real-world environments
