Reinforcement Learning, Deep Learning’s Partner
This year, we have seen all the hype around AI Deep Learning. With recent innovations, deep learning demonstrated its usefulness in performing tasks such as image recognition, voice recognition, price forecasting, across many industries. It’s easy to overestimate deep learning’s capabilities and pretend it’s the magic bullet that will allow AI to obtain General Intelligence. In truth, we are still far away from that. However, deep learning has a relatively unknown partner: Reinforcement Learning. As AI researchers venture into the areas of Meta-Learning, attempting to give AI learning capabilities, in conjunction with deep learning, reinforcement learning will play a crucial role.
What is Reinforcement Learning?
Imagine a child who is learning by interacting with their environment. Each touch will generate a sensation that can result in a reward. For instance, the pleasant smell of the flower will entice the child to want to smell the flower again; the pain from a prick of the flower’s stem will alert the child who will refrain from touching the stem again.
In each case, as the child interacts with the environment, the environment reciprocates and teaches the child by rewarding the child with different sensations.
The child is learning by trial and error.
This is reinforcement learning. In reinforcement learning, an agent starts in a neutral state. Then, as actions are taken, the environment helps the agent transition from the neutral state to other states. In these other states, there might be rewards for the agent.
The goal of the agent is to gather as many rewards as possible.
You can visualize yourself as an agent, walking on a reinforcement learning path, starting at the beginning of a maze toward the exit. With each step that you take, you have a chance of collecting rewards that you can tally up. Depending on the type of rewards and the quantity of the rewards, in your reward pouch, decisions can be made to direct you toward the exit. Eventually, with many tries, an optimal path can be found through the maze.
Applications of Reinforcement Learning
Reinforcement learning is already widely used in many industries such as manufacturing, inventory management, delivery management, and finance.
In manufacturing, robots use reinforcement learning to learn specific tasks in sequence with precision on the assembly line.
In delivery management systems, reinforcement learning can be used to split the customer’s order among different vehicles on different routes to arrive at the destination.
In the financial industry, reinforcement learning is used to evaluate trading strategies to fulfill financial objectives.
Most of the time, these reinforcement learning algorithms are integrated with deep learning algorithms to create deep reinforcement learning algorithms that can handle more complex tasks.
Why is Deep Learning Limited?
In contrast, deep learning is a subset of machine learning that’s receiving a lot more attention than reinforcement learning. People often think that deep learning is the part of machine learning that will ultimately bring us closer to Artificial General Intelligence.
In truth, deep learning is only part of the solution.
Deep learning is based on a neural network architecture that’s similar to the human brain to allow it to make simple decisions.
Traditional models of deep learning such as convolutional neural networks or recurrent neural networks are great at making simple decisions and finding hidden relationships. However, due to the restrictions in the number of layers of complexity, time, and data, deep learning is limited in its capacity to make decisions in complex situations.
On top of that, deep learning by itself cannot infer the meaning of the patterns that it sees. Most technologists will agree that deep learning, by itself is great for classification problems, but it is inadequate for problems that require reasoning, understanding and common sense.
Deep learning also has a huge limitation: it needs to consume a lot of data to be able to make accurate decisions. The biggest asset in human decision making is that our brains can zoom in on “critical” information. Often, we make important decisions on little current information in conjunction with our past experiences and knowledge from the past.
Artificial General Intelligence can only become more sophisticated if it can also make decisions based on small amounts of data.
In other words, deep learning by itself does not bring us to Artificial General Intelligence. It needs a partner that can perform abstractions and reasoning.
Reinforcement Learning’s Role in Artificial General Intelligence
One of the most important aspects of research into Artificial General Intelligence is in the area of reinforcement learning. While Deep Learning gives AGI the ability to uncover hidden patterns to make connections, it is reinforcement learning that allows AGI to make abstractions to understand the meaning behind patterns and in turn direct behavior.
When deep learning is combined with reinforcement learning in deep reinforcement learning, the AGI can plan, understand, and strategize the actions that it should take.
An example of this is DeepMind’s MuZero algorithm, a deep reinforcement learning algorithm that’s able to construct agents that can plan out how to play games such as chess and GO, without knowing the rules.
This is the first step of artificial general intelligence.
When AI can use limited data to plan, understand and strategize actions, without being explicitly “taught”, then the AI is closer to achieving general intelligence.
In the coming year, we will likely see a lot more applications of deep reinforcement learning algorithms in different industries. This is an exciting time. With many applications in different industries, the usage of these algorithms will become more prevalent and sophisticated. With many iterations of research and application, we can truly see the potential of power of AI that might someday achieve general intelligence.