Reinforcement learning is a machine learning training method based on rewarding desired behaviors and/or punishing undesired ones. In general, a reinforcement learning agent is able to perceive and interpret its environment, take actions and learn through trial and error.
In reinforcement learning, developers devise a method of rewarding desired behaviors and punishing negative behaviors. This method assigns positive values to the desired actions to encourage the agent and negative values to undesired behaviors. This programs the agent to seek long-term and maximum overall reward to achieve an optimal solution.Content Continues Below
These long-term goals help prevent the agent from stalling on lesser goals. With time, the agent learns to avoid the negative and seek the positive. This learning method has been adopted in artificial intelligence (AI) as a way of directing unsupervised machine learning through rewards and penalties.
Common applications of reinforcement learning
While reinforcement learning has been a topic of much interest in the field of AI, its widespread, real-world adoption and application remain limited. Noting this, however, research papers abound on theoretical applications, and there have been some successful use cases.
This article is part of
Current use cases include, but are not limited to, the following:
- resource management
- personalized recommendations
Gaming is likely the most common usage field for reinforcement learning. It is capable of achieving superhuman performance in numerous games. A common example involves the game Pac-Man.
A learning algorithm playing Pac-Man might have the ability to move in one of four possible directions, barring obstruction. From pixel data, an agent might be given a numeric reward for the result of a unit of travel: 0 for empty space, 1 for pellets, 2 for fruit, 3 for power pellets, 4 for ghost post-power pellets, 5 for collecting all pellets and completing a level, and a 5-point deduction for collision with a ghost. The agent starts from randomized play and moves to more sophisticated play, learning the goal of getting all pellets to complete the level. Given time, an agent might even learn tactics like conserving power pellets until needed for self-defense.
Reinforcement learning can operate in a situation as long as a clear reward can be applied. In enterprise resource management (ERM), reinforcement learning algorithms can allocate limited resources to different tasks as long as there is an overall goal it is trying to achieve. A goal in this circumstance would be to save time or conserve resources.
In robotics, reinforcement learning has found its way into limited tests. This type of machine learning can provide robots with the ability to learn tasks a human teacher cannot demonstrate, to adapt a learned skill to a new task or to achieve optimization despite a lack of analytic formulation available.
Reinforcement learning is also used in operations research, information theory, game theory, control theory, simulation-based optimization, multiagent systems, swarm intelligence, statistics and genetic algorithms.
Challenges of applying reinforcement learning
Reinforcement learning, while high in potential, can be difficult to deploy and remains limited in its application. One of the barriers for deployment of this type of machine learning is its reliance on exploration of the environment.
For example, if you were to deploy a robot that was reliant on reinforcement learning to navigate a complex physical environment, it will seek new states and take different actions as it moves. It is difficult to consistently take the best actions in a real-world environment, however, because of how frequently the environment changes.
The time required to ensure the learning is done properly through this method can limit its usefulness and be intensive on computing resources. As the training environment grows more complex, so too do demands on time and compute resources.
Supervised learning can deliver faster, more efficient results than reinforcement learning to companies if the proper amount of data is available, as it can be employed with fewer resources.
How is reinforcement learning different from supervised and unsupervised learning?
Reinforcement learning is considered its own branch of machine learning, though it does have some similarities to other types of machine learning, which break down into the following four domains:
- Supervised learning. In supervised learning, algorithms train on a body of labeled data. Supervised learning algorithms can only learn attributes that are specified in the data set. Common applications of supervised learning are image recognition models. These models receive a set of labeled images and learn to distinguish common attributes of predefined forms.
- Unsupervised learning. In unsupervised learning, developers turn algorithms loose on fully unlabeled data. The algorithm learns by cataloging its own observations about data features without being told what to look for.
- Semisupervised learning. This method takes a middle-ground approach. Developers enter a relatively small set of labeled training data, as well as a larger corpus of unlabeled data. The algorithm is then instructed to extrapolate what it learns from the labeled data to the unlabeled data and draw conclusions from the set as a whole.
- Reinforcement learning. This takes a different approach altogether. It situates an agent in an environment with clear parameters defining beneficial activity and nonbeneficial activity and an overarching endgame to reach. It is similar in some ways to supervised learning in that developers must give algorithms clearly specified goals and define rewards and punishments. This means the level of explicit programming required is greater than in unsupervised learning. But, once these parameters are set, the algorithm operates on its own, making it much more self-directed than supervised learning algorithms. For this reason, people sometimes refer to reinforcement learning as a branch of semisupervised learning, but in truth, it is most often acknowledged as its own type of machine learning.
Types of reinforcement learning algorithms
Rather than referring to a specific algorithm, the field of reinforcement learning is made up of several algorithms that take somewhat different approaches. The differences are mainly due to their strategies for exploring their environments.
- State-action-reward-state-action (SARSA). This reinforcement learning algorithm starts by giving the agent what's known as a policy. The policy is essentially a probability that tells it the odds of certain actions resulting in rewards, or beneficial states.
- Q-learning. This approach to reinforcement learning takes the opposite approach. The agent receives no policy, meaning its exploration of its environment is more self-directed.
- Deep Q networks. These algorithms utilize neural networks in addition to reinforcement learning techniques. They utilize the self-directed environment exploration of reinforcement learning. Future actions are based on a random sample of past beneficial actions learned by the neural network.
Some of the main ideas of reinforcement learning -- in particular, the idea of trial-and-error learning -- date as far back as the mid-1950s.
Computing pioneer Marvin Minsky and his paper, "A Neural-Analogue Calculator Based upon a Probability Model of Reinforcement," was critical during this early period. This 1952 paper was among the first to describe a self-learning algorithm that learns by rewarding positive actions.
At the time, however, supervised and reinforcement learning were not clearly separated terms. The resulting confusion meant papers done with reinforcement learning in mind were instead based around supervised learning throughout the 1960s and 1970s. However, in 1963, Donald Michie published a paper describing a machine that successfully deployed a type of trial-and-error learning to play the game tic-tac-toe.
Reinforcement learning started to become more formalized in the 1980s, mainly through the work of A. Harry Klopf, who focused on adding a reward-punishment component to an agent's trial-and-error learning. Richard S. Sutton and Andrew G. Barto's work that followed differentiated supervised and reinforcement learning and built upon their predecessors work on the latter.