Products / BakingSoDA

Reinforcement Learning

The Concept of Reinforcement Learning

Reinforcement Learning is widely used as the most effective method for optimization analysis among various machine learning techniques. Reinforcement learning is a method of self-learning decisions to maximize learning outcomes. Unlike supervised learning, another technique of machine learning, it uses rewards instead of targets and executes policies or action instead of producing predicted values. Optimized design, such as State, Reward, Environment, and Action, suitable for each business situation becomes important in reinforcement learning implementation. 

Reinforcement learning: self-learning decision making that maximizes future reward value based on the result of decision making in the absence of sufficient historical data to learn

Ex) Learning the bank loan execution strategy by itself to find the optimal execution strategy

Unsupervised learning: Analysis methods that find similar groups or patterns using only data without predicted target values

Ex) Cluster analysis, association analysis, network analysis, etc.

Supervised learning: Analysis methods that build a predictive model that fits the results when past execution data is sufficient to learn

Ex) Classification that calculates scores or regression

Definition of Reinforcement Learning

Reinforcement Learning is one of the various machine learning methods, in which an agent defined in an environment recognizes the current state and selects an action or sequence of actions to maximize reward among the selectable actions.

Agent: A goal-oriented subject that observes state and selects actions

Environment: Problem setting encountered by agents
State: The current situation of the agent

Action: The agent's action in the current state

Reward: Information that informs the good or bad of the result of an action

Interaction of Reinforcement Learning

Agent observes self status in environment.

Choose an action based on a certain criterion (i.e. value function: the sum of rewards you expect to receive in the future by taking this action)

Execute the selected action in the environment
Transfer to the following state and get rewards from the environment
Revise agent's information through rewards

Applications of Reinforcement Learning

Reinforcement learning is commonly used for games or robotics. In addition to games and robotics, AgileSoda defines problems and environments that can be solved through reinforcement learning in accordance with the business environment of the company, and conducts researches to help companies solve problems on their own by applying algorithms (Policy Gradient, etc.) suitable to each companys business circumstance. In particular, as data accumulates, we plan to develop functions and models that can improve and recommend strategies by themselves, and will soon install those on SparklingSoDA. Currently, consulting services for enterprise clients are provided, please contact your sales representative for details.

Generative Adversarial Network (GAN)

Defination of GAN

GAN represents 

G   Generative

A   Adversarial

N   Network

In other words, GAN is a model created by using adversarial networks in generating data.

GAN is divided by two parts. 


  This is the part that creates data.

  You can understand it as a counterfeiter.

  It is trying to trick Discriminator


  This is the part that distinguishes whether data is real or fake.

  You can understand it as a police officer.

  It distinguishes whether the data created by Generator is fake or real.

Finally Generator studies and generates data which Discriminator(Policeman) can't figure out and is deceived by Generator(Money counterfeiters). 

More details will be found on the homepage below. Deep Learning for Computer Vision: Generative models and adversarial training (UPC 2016)


GAN has evolved a lot in the direction of generating image data, and recently, there have been attempts to generate words through GAN. Unlike these attempts, We are conducting research to improve the performance of the model by generating the data necessary for modeling based on the actual corporate data through GAN and Embedding. We will provide a breakthrough solution in situations where it is difficult for enterprises to adopt or operate machine learning due to missing data or the lack of data to learn.

We call it Data Aesthetic service.

Like skin regeneration / like skin whitening / data automatic (reproduction) / data noise removal / data embedding / Like face reduction


Embedding Definition

Embedding is a way of representing data, mapping individual objects, such as words, into vectors of real numbers. The way of expressing individual objects such as words in a simple way is called local representation, and the way of expressing objects as a vector of real numbers is called distributed expression. Embedding is a technique that expresses localized data in a distributed representation manner. Typical algorithms include Word2Vec and Doc2Vec.

Word2Vec is an algorithm that embeds a word into a vector of real numbers, and Doc2Vec is an algorithm that embeds a single document (such as news articles and blog posts) into a vector of real numbers.

More details would be found on homepage.


Word2Vec & Doc2Vec


AgileSoDA supports companies to transform AI capabilities and output into internal intellectual assets and to resolve data quality problems. It also provides algorithm-based technologies and services including reinforcement learnings for optimal business decision.

Contact US