Reinforcement Learning In A Nutshell

Reinforcement learning (RL) is a subset of machine learning where an AI-driven system (often referred to as an agent) learns via trial and error.

Understanding reinforcement learning

Reinforcement learning is a technique in machine learning where an agent can learn in an interactive environment from trial and error. In essence, the agent learns from its mistakes based on feedback from its own actions and experiences.

Reinforcement learning is similar to supervised learning in that both approaches map an input variable to an output variable. Unlike supervised learning, which provides feedback in the form of a correct set of actions, reinforcement learning uses rewards and punishments as feedback for positive and negative behavior.

To understand why an agent would be subject to rewards and punishments, note that the objective of reinforcement learning is to discover an action model that maximizes the total cumulative reward of the agent.

In the context of the current AI paradigm, reinforcement learning from human feedback enables a large language model to become much more specialized.

And at the same time, the same process can be used to make it less biased. Now, the interesting take here is that today’s network effects for AI players are built via reinforcement learning feedback loops.

In short, with reinforcement learning, an AI agent learns to make decisions by performing actions in an environment and receiving feedback as rewards or penalties.

OpenAI shared already back in 2017 how this process was instrumental in developing safe AI systems. And yet this same methodology proved quite effective also to make these AI systems way more effective for specific tasks.

To be sure, reinforcement learning from human feedback wasn’t a discovery of OpenAI but an achievement of academia.

Yet, what the OpenAI team was good at, was in scaling this approach.

Back then, the team at OpenAI trained an algorithm with 900 bits of feedback from a human evaluator to make it learn to backflip.

Of course, that doesn’t seem a huge achievement for a simple and narrow task, and yet, this was the embryonic stage of what would later turn into something like ChatGPT.

Reinforcement learning differs from supervised and unsupervised learning in that the model is not trained on labeled data but instead learns from its interactions with the environment.

The process involves the agent observing the state of the environment, taking action, and receiving a reward signal based on the outcome of its actions.

Positive and negative reinforcement in RL

What constitutes positive and negative reinforcement, exactly? Let’s have a look.

Positive reinforcement

Positive reinforcement is an event that occurs in response to a behavior that increases its frequency and strength. That is, when the agent performs the correct action, it receives positive feedback or a positive reward.

Positive reinforcement maximizes agent performance and sustains change for a longer period. It is thus the most common type of reinforcement used.

Negative reinforcement

In the context of training a model, negative reinforcement is used to maintain a minimum performance standard as opposed to enabling the model to maximize its performance.

Negative reinforcement is used to keep the model away from undesirable action. However, this approach does not encourage the model to seek out more desirable actions.

The basic elements of reinforcement learning

Reinforcement learning can be illustrated with a simple diagram that demonstrates the action-reward feedback loop. The diagram contains the following annotations and key terms:

Environment – the world in which the agent lives, interacts, and receives feedback.
Action – the set of all moves an agent can potentially make.
Reward – feedback from the environment for actions that lead to a successful state.
State – the current situation of the agent in their environment. It can be a specific moment or a specific position.
Policy – the policy defines the strategy the agent will use to pursue its objectives based on the current state. The agent maps actions to states to determine which action has the highest reward, and
Value function – the reward an agent would receive if it undertook an action in a particular state. In other words, how favorable is a certain state for the agent?

Reinforcement learning applications

To conclude, we’ve detailed two examples of how reinforcement learning is applied in the real world.

Robotics

RL is used in robotics to create adaptive control systems that learn from their own behavior experiences.

There is also promise that the technique can overcome the curse of dimensionality, a problem robots experience in three-dimensional environments where they have less data to make decisions as the volume of the space increases.

Industrial automation

Industrial automation is another application with potential.

DeepMind has used reinforcement learning technologies to help Google reduce the energy consumption of heating, ventilation, and air conditioning (HVAC) in its data centers.

Microsoft’s Bonsai is another project that offers low-code, AI-powered automation to improve efficiency, reduce downtime, and optimize process variables. One example is the use of artificial intelligence to replace skilled human operators on tuning machines and other equipment.

Key takeaways

Reinforcement learning (RL) is a subset of machine learning where an AI-driven system (often referred to as an agent) learns via trial and error.
Unlike supervised learning, which provides feedback in the form of a correct set of actions, reinforcement learning uses rewards and punishments as feedback for positive and negative behavior.
Two of the major applications of reinforcement learning are robotics and automation. In the case of the latter, it is seen as an effective way to reduce operational inefficiencies and downtime.