How Psychology Drives Machine Learning

Reinforcement Learning and Conditioning

People have continuously wanted to create machines that can think, learn, and reason. The research within the field of artificial intelligence (AI) leads us to believe that we should look at algorithms, thinking they are comparable to our human ways of thinking and reasoning. A good example is the following definition of AI as “any machine that does things a brain can do.”

This consideration also applies to reinforcement learning: It’s a machine learning field that accroding to Prakriteswar Santikary is concerned with how software agents should take actions in an environment to maximize rewards. It’s one of the three paradigms in machine learning and builds on a well-known educational/psychological concept: conditioning.

Conditioning Methods

Pavlov’s behavioral learning theory (also known as classical conditioning) states that a new, conditional reaction can be added to a natural, mostly innate, so-called unconditional response through learning. A well-known example is a Pavlovian dog: every time the dog detected food (the biological stimulus), a bell would ring (neutral stimulus). After a few such feedings, the dog would react to the bell’s sound in the same way it would respond to food: its salvia began to flow. Ivan Pavlov first studied this in 1897 and to identify other types of classical conditioning: forward conditioning, simultaneous conditioning, backward conditioning, and temporal conditioning.

Based on this theory, other learning methods, such as operant conditioning, also known as learning by success, were developed. These paradigms of behavioristic learning psychology concern the understanding of stimulus-response patterns from originally spontaneous behavior. The frequency of action is changed permanently by its pleasant (appetitive) or unpleasant (aversive) consequences. This means that desirable behavior is reinforced by reward, and undesirable behavior is suppressed by punishment.

We either award the agent (reinforcement) or punish the agent for unwanted behavior (punishment). And based on that, there are two types of reinforcement learning methods:

• Positive: It is characterized as an occasion that happens because of a particular behavior. It increases the quality and the recurrence of the behavior and impacts emphatically on the activity taken by the agent.
• Negative: This scenario centers around the strengthening of action that occurs because of an adverse condition that has to stop or be avoided.

Reinforcement Learning for Neural Networks

As a neural network learning method, reinforcement learning algorithms should help to attain a complex objective or maximize a particular measurement over numerous steps. Typically, reinforcement learning methods are applied in motion control scenarios, creating training systems that provide instructions and materials according to requirements. This method can also be used, as distinct from supervised learning, when there is insufficient data.

Reinforcement learning is applied in robotics for industrial automation and business strategy planning. It can be assumed that reinforcement learning algorithms perform better over time in more ambiguous, realistic environments when selecting an arbitrary number of possible actions. Some even believe that this algorithm is the most promising path to solve complex problems around strong AI, given enough data and calculations are available.

There is one significant downside to the concept of conditioning: The study of learning through conditioning is strictly limited to observable behavior and does not speculate on constructs that may underlie the behavior. Therefore, it does not clarify how learning by intrinsic motivation (e.g., curiosity) works.

The same is true for algorithms because they are rule-based, not rule-bound.