Machine Learning: Achieving Ultimate Intelligence
Machine learning techniques may be applied to solve a whole range of problems, and today we take a look at some of the more prevalent examples out there.
By Tetiana Gladkikh , Taras Hnot and Volodymyr Solskyy
It is not a secret that our success is directly connected to our ability to make right decisions applying our knowledge and experience shaped by different factors. The reason why it is sometimes so difficult to come to decision lies within the existing conditions and possible choices, variability of which leads to an interesting paradox:
- Knowledge growth boosts decision-making efficiency.
- The less time you have to process information, the less effective your solution will be.
Extracting useful information from available data is the main task of machine learning.
Pattern Recognition and Anomaly Detection
Occupying a special place in machine learning, decision-making is viewed as an attempt to find the most interesting patterns in the processes with the subsequent results interpretation.
This problems category relates to machine learning problems subclass - Pattern Recognition, where decision-making rides on learning without a teacher. For example, detecting patterns in data of a particular organization leads to better understanding of its internal processes and thus detecting anomalies in its network activities (such as data breaches, intrusions, internal threats, virus activity, etc.). Detected anomalies, or in other words any activities deviating from what is considered a norm, may be split into three types:
- Dynamic Threshold Model: significant deviation of the observed values.
- Association Rules Based Model: unusual set of the observed values of the measured parameters.
- Time Series Clustering Model: unusual dynamics in the observed process.
Since all three models have their pros and cons of implementation, merging them into one ensemble helps neutralize their disadvantages and adapt to new or changing conditions.
Bitcoin Network
Another example of machine learning techniques for community structure analysis is a piece of our Data Science Group work related to Bitcoin network. Having analyzed Bitcoin users activity, their wallets features, and connections while taking into account known groups, Bitcoin events, and general trends, we aimed to uncover behavior patterns from large data sets; this case of machine learning technique helped us better understand the hidden, underlying principles guiding the development of the Bitcoin network. The following animation presents network evolution over time.
Fig. 3. – Bitcoin Graph development
[caption id="attachment_16675" align="aligncenter" width="763"]
Fig. 4. – Bitcoin graph with communities’ visualization (gambling centers and mining pools)[/caption]
Moreover, we described a way to Bitcoin wallets deanonimization and showed the most significant patterns of coins flow that can be used as a basis for abnormal cases detection.
Intelligent Agents and Deep Reinforcement Learning
One more example of using unsupervised machine learning approach is development of intelligent agents whose behavior can adapt to changes in the external environment. Solution to this problem is based on models with extremely complex architecture, which presupposes application of deep learning technology. Again, our Data Science Group implemented a deep reinforcement learning algorithm described in Playing Atari with Deep Reinforcement Learning paper by DeepMind.
[embed]https://www.youtube.com/watch?v=T58HkwX-OuI[/embed]
Fig.5. – Deep Reinforcement Learning - video
The developed model imitates human playing and is based on the results of human interaction with the external environment. The experiment was extended by the results of a comparative analysis of the human and training model (agent) impact under different variations of the game. The graph below shows the results of the Breakout game analysis as played by a trained agent and two real players for the 30 grades of the delay. Selection of the delay is based on the results of the human response study.
[caption id="attachment_16685" align="aligncenter" width="619"]
Fig. 6. – Performance comparison[/caption]
As you can see, the trained agent shows way better results when it comes to an extremely high speed of reaction once all the inputs are unified. This suggests a possibility of using this kind of model for building adaptive control systems such as:
- Automatic process control
- Intelligent video monitoring
- Intelligent agents to carry out operations in the inaccessible areas
Conclusions
Machine learning techniques may be applied to solve a whole range of problems, and today there hardly is a single industry that wouldn’t benefit from improved decision making brought by identifying hidden patterns in data with their subsequent formalization. If machine learning is not on your radar yet, make sure you’re getting closer to it!
About the Authors
Tetiana Gladkikh is a Data Scientist at SoftServe. She has 19 years of experience in Research, with 15 years being dedicated to Data Mining and Computational Intelligence. The main areas of Tetiana's scientific interests are Data Mining, Artificial Intelligence (Genetic Algorithms, Neural Network, Fuzzy Logic), Mathematical Statistics, and Computer Vision.
Taras Hnot is a Data Analyst at SoftServe. The main areas of his interests are Statistical Learning, Predictive Analytics, Time Series Analysis, Artificial Intelligence and Recommender Systems. Taras has experience in development of anomaly detection systems, analyzing and detecting patterns of huge payment networks, implementing different types of algorithms in order to build computer vision systems.
Volodymyr Solskyy is a Data Scientist at SoftServe. He has 9 years' experience in commercial software development, being immerced for 5 years in Scalable Architectures, Cloud Technologies and Machine Learning. His main areas of interest are Distributed Systems, Network Analysis, Information Theory and Knowledge Extraction.