Bayesian Deep Multi-Agent Multimodal Reinforcement Learning for Embedded Systems in Games, Natural Language Processing and Robotics

Abstract

Nowadays, Machine Learning is one of the most dynamic fields, as it attracts strong research interest from both industry and academia alike. It is not surprising that a huge amount of funding from government agencies, universities, Tech giants and well-funded startups is currently being allocated exclusively to this field. Reinforcement Learning, one of the three major subfields of Machine Learning, has recently gained a tremendous traction due to the fact that algorithms can run more efficiently. This is mainly due to two reasons, firstly affordable and portable hardware, such as mobile phones, wearables and Internet of Things devices, now has the capacity to run these algorithms and secondly, new methods and models are being proposed that deal with the matter of efficiency from an algorithmic point of view. This proposal is concerned with dealing with open challenges in memory efficiency, while devising and applying such Reinforcement Learning algorithms for embedded systems in the domains of Games, Natural Language Processing and Robotics using Deep Learning models and Bayesian inference, a very powerful framework. Natural Language Processing is a domain of Machine Learning in which the input is given in the form of a text from a natural language that human agents use for everyday communication. It is also a domain that still faces a series of ongoing challenges, as opposed to more saturated domains, such as Computer Vision. One of them is the fact that ground truth is difficult to be decided due to the nature of text in general. Other challenges include the personalized type and tone of the conversation held by the human agents, such as formal, informal, aggressive, polite, etc. Therefore, this proposal deals with all of these matters, mainly in the subdomain of Question Answering systems, also known as chatbots, in a multi-agent setting.

An Overview of the GUT-AI Foundation: Vision for an Ecosystem of Concepts and Implementations

“…In this case, determinism oscillates between the two states of Red Queen dynamics and White Queen dynamics, similar to a pendulum oscillating between two extrema. Even within a single state, such as the Red Queen dynamics, there are oscillations between entities (e.g., host-parasite co-evolution), which is compatible with the pyramidal fractal pattern [1], also proposed by GUT-AI theory. As a direct consequence of such a pattern, the theory also explains how physical constraints or limits (e.g., Planck length, Planck time, Shannon limit, speed of light) [1], such as the physical size of an entity can and will affect the expression of the two aforementioned states.…”

“…Even within a single state, such as the Red Queen dynamics, there are oscillations between entities (e.g., host-parasite co-evolution), which is compatible with the pyramidal fractal pattern [1], also proposed by GUT-AI theory. As a direct consequence of such a pattern, the theory also explains how physical constraints or limits (e.g., Planck length, Planck time, Shannon limit, speed of light) [1], such as the physical size of an entity can and will affect the expression of the two aforementioned states. For instance, homogeneous entities (e.g., unicellular organisms, generic AI solutions or homogeneous RL agents) exhibit more White Queen dynamics than Red Queen dynamics, whereas heterogeneous entities (e.g., multicellular organisms, industry-specific AI solutions or heterogeneous RL agents) exhibit relatively more Red Queen dynamics than White Queen dynamics.…”

“…Even relatively small variations (or oscillations) could create immeasurable problems for measuring devices, critical infrastructure, life-saving equipment, etc. For example, Sheldrake [34] proposes 'habits of nature' (which can be inherited by species), which is compatible with and could be a hypothesis for the emergence of 'learnable rules' as proposed by the GUT-AI theory [1],…”

“…If an Independent Researcher dares to overstep to a field next to him, then he should prepare himself for a battle! As also explained in [1], this is in direct contrast to Ancient Greeks who were not confined to a particular specialization or interest group, but were polymaths. This kind of enmity creates all sorts of problems, since phenomena around us can rarely be interpreted as a mere collection of independent components (i.e.…”

“…Due to the fact that this is a hypothesis that describes an ecosystem and the entities interacting with it and also among themselves (i.e. a Complex System [1,36]), only a multidisciplinary approach can be used to both mathematically and empirically study this proposed phenomenon, due to the fact that it necessarily requires knowledge from various disciplines, including Game Theory, Network Science, Optimization, Control Theory, Dynamical Systems, Collective Intelligence, etc. percentage of 'oscillatory' determinism (e.g., 80%), both of which need to be 'learnable'.…”