A new area of AI research, called Deep reinforcement learning (DRL) is making waves within artificial general intelligence (or AGI) circles, writes Sam Charrington for Venture Beat. This is because DRL mirrors human learning by exploring and receiving feedback from environments.
Supervised machine learning trains models based on “known-correct” answers. By contrast, researchers implement reinforcement learning by having an “agent” interact with an environment. Thus, when the agent’s actions produce a desired result, it receives positive feedback. The promise of DRL has led to a number of startups hoping to capitalize on this technology, writes Charrington. Pieter Abbeel at the University of California Berkeley has founded Embodied Intelligence, a startup that will combine VR (virtual reality) and DRL and apply it to robotics. Pit.ai plans to out-trade traditional hedge funds by applying it to algorithms.
Increased interest has also led to startups creating new open source toolkits and environments for training DRL agents. Charrington lists several of these interfaces, like House3D (a collaboration between UC Berkeley and Facebook AI researchers). It offers over 45,000 simulated indoor scenes with realistic room and furniture layouts. The primary task, introduced in the team’s paper, is “concept-driven navigation,” which would train an agent to navigate to rooms within a house.
Charrington cautions that applying DRL to non-trivial problems creates the challenge of constructing a reward function that encourages desired behaviors without the adverse effect of promoting cheating. However, with all the tools and platforms in development, researchers hope to eventually work through these challenges.