Citation:
Abstract:
To generate adaptive behaviors, animals must learn from their interactions with the environment. Describing the algorithms that govern this learning process and how they are implemented in the brain is a major goal of neuroscience. Careful and controlled observations of animal learning by Thorndike, Pavlov and others, now more than a century ago, identified intuitive rules by which animals (including humans) can learn from their experiences by associating sensory stimuli and motor actions with rewards. But going from explaining learning in simple paradigms to deciphering how complex problems are solved in rich and dynamic environments has proven difficult (Figure 1). Recently, this effort has received help from computer scientists and engineers hoping to emulate intelligent adaptive behaviors in machines. Inspired by the animal behavior literature, pioneers in artificial intelligence developed a rigorous and mathematically principled framework within which reward-based learning can be formalized and studied. Not only has the field of reinforcement learning become a boon to machine learning and artificial intelligence, it has also provided a theoretical foundation for biologists interested in deciphering how the brain implements reinforcement learning algorithms. The ability of reinforcement learning agents to solve complex, high-dimensional learning problems has been dramatically enhanced by using deep neural networks (deep reinforcement learning, Figure 1). Indeed, aided by ever-increasing computational resources, deep reinforcement learning algorithms can now outperform human experts on a host of well-defined complex tasks …