Agents are becoming more artificially intelligent but a number of problems must be solved before we can unleash them into our physical world (i.e. robots) and trust them in our digital one.
At the recent NAVER Cloud AI Summit, Research Fellow Christopher Dance presented two research challenges that must be addressed; ensuring that intelligent agents act in a risk-averse fashion and how to have high-level control over what agents generate. Acting risk-averse and making a special effort to avoid relatively rare but serious outcomes is of central importance for robots, drones and vehicles that move around in the presence of humans. Having high-level control will help us to explain to robots what we want them to do. It can also guarantee that key messages are conveyed in text that is generated by agents as well as limit bias that may lie within.
Keywords: risk-averse reinforcement learning, inverse reinforcement learning and imitation, hierarchical reinforcement learning, meta-learning, universal policies, natural language generation