Deep Reinforcement Learning: Future Frontiers of Artificial Intelligence
Artificial Intelligence, though from a marketing perspective of different organizations may mean a lot of things encompassing systems ranging from conventional analytics, to more contemporary deep learning and chatbots. But technically the use of Artificial Intelligence (AI) terminology is restricted to the study and design of ‘Rational’ agents, which could act ‘Humanly’. Of the many definitions given by different researchers and authors of AI, the criteria for calling an agent an AI agent is that it should possess ability to demonstrate ‘thought process and reasoning’, ‘intelligent behavior’, ‘success in terms of human performance’, and ‘rationality’. This identification should be our guiding factor to identify the marketing jargons from real AI systems and applications.
Among the different AI agents, Reinforcement Learning agents are considered amongst the most advanced and capable of demonstrating high level of intelligence and rational behavior. A reinforcement learning agent interacts with its environment, which could demonstrate multiple states, and acts to change the environment state, thereby also receiving a reward or penalty as determined by the state and the objective of the agent. This definition may look simple, but it led to the development of a lot of advanced AI agents to perform many complex tasks, sometimes challenging human performance in specific tasks.
But to really challenge human capabilities in most of the tasks that we could accomplish, needs the processing of a lot of different states, which may require different senses as humans possess. Also the migration from one state to another in real-life scenario may be quite complex. Assume the difference between self-driving cars and a human driver. To reach human level proficiency in driving in a real life scenario, the agent needs to have capability to work on complex states comprising of a combination of visuals coming from many high definition cameras, real time high frequency data from hundreds of sensors, and live server and GPS feeds. This task to work on a fusion of such complex and disparate states types with complex and non-deterministic state transitions is not trivial, and may require an enhancement with such sub-systems that may process all these different types of senses and data.
This is where Deep Learning excels. Deep learning uses similar neurons in different architectures to emulate different senses that we possess and more. Ranging from simple classification of structured data, to image processing, object detection, instance segmentation, speech recognition and generation, to Natural Language processing Deep Learning (DL) is in the forefront of advancements in all these fields. Though DL can provide rich insights from these different types of structured and unstructured data, but DL on its own cannot act like an agent and use its own insights to optimise a given goal. But when we combine Deep Learning with Reinforcement, we enter the realm of the most powerful AI agents based on Deep Reinforcement Learning.
Deep Reinforcement Learning (DRL) has recently shown a lot of promising advancements and has exceled against human intelligence in multiple fields. Besides self-driving cars, DRL has been used to defeat humans in most of the games ranging from our nostalgic Atari and Mario like games to some of the most complex games like Alpha-Go. DRL is in its very nascent state and most of the research work that led to these advancements is very recent. But given what it has already accomplished, and with more active research and investments that it is attracting, we can safely assume that it is the Future Frontier of Artificial Intelligence, using which the AI agents will surpass most of human abilities by many folds, and also give rise to new fields of applications which does not exists today due to limitations of evolution of human intelligence.
by Mohit Sewak