SouCI: Socially Sustainable Computational Intelligence.

点击翻译按钮获取中文摘要

My research aims to build socially sustainable computational intelligence (SouCI). SouCI combines a view that the way humans learn is fundamentally hierarchical and generative/predictive, with a practical model-based hierarchical Bayesian reinforcement learning model, the partially observable Markov decision processes, or POMDP. POMDPs are viewed as general purpose models for autonomous agent policy computation in uncertain, yet reasonably predictable environments. Further, by making the parameters of these models part of the model itself, an agent learns how the world works, and what its role is, at the same time. Model parameter spaces are continuous, however, posing a significant computational challenge. My aim is to overcome this challenge by leveraging insights from emotional processing in the human brain. It is now known that emotions play a significant role in helping people overcome the computational difficulties posed by action in an uncertain, particularly social, world. Emotions are thought to serve as "somatic markers" that allow people to analogically transfer skills used in one domain to another. I plan to use the same types of mechanisms to more efficiently learn these hybrid, emotion-based, POMDPs. I aim to build computational intelligence that mirrors the social and emotional aspects of human intelligence, helping pave the way towards autonomous agents and robots that are genuine members of a society pervaded by humans and robots which aims for sustainability of natural resources (environmental), society (economic), and freedom, justice and equality for people (social). I will build a social learning system called SouCI that aims for social sustainability across a range of application areas, and will show how these can lead to economic growth and environmental stability as well. SouCI may be divided into three interrelated components: learnability, dilemma resolution, and ethics/fairness. Agents must be capable of distributed learning of shared policies of action, should they wish to move away from the trap of pure costly rational decision making. Further, these mechanisms must be such as to enhance cooperation, such as to provide a solution to social dilemmas underlying sustainability (e.g. free-riding). Such agents should be "programmable'' in the sense that one could explicitly program a measure of fairness into the high level predictive (generative) models used by the agent. Such agents are also required to learn models which aim for peace, diversity, inclusion, progress, and equity, rather than the opposites. The degree to which an agent trades-off these different elements is then the "tuning knob'' that can be used in prior model specification to center an agent's ethical reasoning around the cultural norm. All three components have the common goal of motivating itself and its group (including humans) to be more cooperative.