1 of 37

MACHINE LEARNING THROUGH NEURAL NETWORKS FOR GAMING

Alessandro, Sestito

2 of 37

AI SYSTEMS ARE COOL��DOES IT MAKE SENSE TO USE FOR GAMING?

3 of 37

MEETING AGENDA

5 MACRO POINTS

Introduction
Multilayer Perceptron
Deep learning
TensorFlow
Neural Network for Gaming

4 of 37

INTRODUCTION

5 of 37

WHAT IS THE AI?

In general, an artificial intelligence system represents a computer system capable of having human characteristics such as reasoning, learning and planning

Used to solve difficult problems
AI science can be seen as a collection of generic algorithms
AI needs data to learn from

6 of 37

WHAT IS A GENERIC ALGORITHM?

A genetic algorithm is defined as such because the logic of the algorithm is constant and does not depend on the problem

The algorithm uses the data to self-generate the solution
The problem shifts to computational strength
it is possible to tune the algorithms to obtain more efficient solutions

Neural networks are among the most powerful generic algorithms

7 of 37

A SIMPLE EXAMPLE

Activate the smartphone screen when we take it out of the pocket

Let's read the accelerometer data
Pre-processing
We look for a mathematical rule that solves the problem

Too complex, let's try to solve with AI

8 of 37

A SIMPLE EXAMPLE

We solve the problem through an AI-oriented approach

Collect the data from the sensors when we want to activate the screen and when it must stay off
Process the data to have a compatible format (matrix like)
Do learning and tuning with a neural network

Problem solved

9 of 37

BIOLOGICAL NEURON

At the base of the neural networks is the biological neuron. the human brain contains approximately 100 billion neurons

Dendrite: connects the neuron to others and serves to receive signals from other neurons
Soma: nucleus of the neuron that decides whether to activate based on the input signals
Axon: it serves to send the signal to other neurons

10 of 37

ARTIFICIAL NEURON

The idea of modeling the artificial neuron came from W.S. McCulloch and W. Pitts in 1943 demonstrating that an infinite network of neurons could model a Touring machine

Receives signals from other neurons
The signals are weighted and combined linearly
The result is passed to an activation function which decides whether to activate the neuron
Also called perceptron

11 of 37

MULTILAYER�PERCEPTRON

12 of 37

MULTILAYER PERCEPTRON

A set of connected artificial neurons is also called Multilayer Perceptron and the most common and used networks are those with 3 levels fully connected

The first level is called input and there is a neuron for each input feature
The central layer is called hidden and contains an arbitrary number of neurons
The last level is called output and serves to obtain the solution of the problem.

13 of 37

MULTILAYER PERCEPTRON

Regression

Classification

14 of 37

FORWARD PROPAGATION

Suppose animal data input and dog or cat as output

The feature vector is passed to the input layer
Weights will be applied, and linearly combined values will be passed to the hidden layer
One output neuron will be activated

What if the output is wrong?

15 of 37

BACK PROPAGATION

It is a technique by which weights are updated backwards to reduce network errors

The weights are updated starting from the output node up to the input nodes
The purpose of this phase is to find the best weights that solve the problem

16 of 37

STOCHASTIC GRADIENT DESCENT

It is the process that underlies the learning of each neuron in the network

A function that measures the error is modeled
We want to minimize di error
The first derivative is calculated based on the parameters
The parameters are translated along the direction of the minimum
Parameters are updated

17 of 37

DEEP LEARNING

18 of 37

WHAT IS DEEP LEARNING

Represents a class of neural networks whose neuronal organization is inspired by the human and animal brain
They are used effectively in various contexts including computer vision, automatic recognition of spoken language, natural language processing, audio recognition and bioinformatics
Can be used in unsupervised learning

19 of 37

CONVOLUTIONAL NETWORKS

Feedforward neural network in which the connectivity pattern between neurons is inspired by the organization of the animal visual cortex
These networks have various applications in image and video recognition, recommendation systems and natural language processing

20 of 37

CONVOLUTIONAL NETWORKS

21 of 37

RECURRING NETWORKS

Is a class of artificial neural network that includes neurons connected in a loop.
This interconnection between layers allows one of the layers to be used as a state memory

Allows to model information-dependent temporal dynamic behavior
Mainly used for unsupervised learning problems

23 of 37

WHAT IS TENSORFLOW

TensorFlow is an open-source software library for machine learning

It is mainly used for the support it provides to neural networks
It is used by Google in voice recognition, Gmail, Google Photos and Search
It is written in Python and C++
It is one of the most powerful neural network tools on the market

24 of 37

EXAMPLE

DOG

AIRPLANE

CAT

TRUCK

25 of 37

NEURAL NETWORK FOR GAMING

26 of 37

WHO ALREADY USES IT

Several companies already use machine learning mechanisms to solve complex problems

PokerStars
William Hill

27 of 37

BOT DETECTION

PokerStars uses advanced Machine Learning systems to solve the bot detection problem

AI can be used to build bots that play automatically for people
It can be difficult to spot bots among players
Using Machine Learning can give better results than traditional approaches

https://www.youtube.com/watch?v=T1XXHC8lXcU

28 of 37

ODDS PREDICTION

William Hill instead uses Machine Learning to solve the Odds Prediction problem

What is the odds price such that the revenue and the number of plays are maximized?
Choose the odds through the match history
Complex problem for traditional approaches

29 of 37

HOW DO THEY DO IT?

The techniques used are not disclosed
They are likely to use standard approaches
Algorithms of this type are worth a lot
Exploit the data they possess
The value of algorithms is intrinsic to the data they possess

30 of 37

WHAT CAN WE DO AS RINGMASTER?

31 of 37

FRAUD DETECTION

It consists of a sequence of plays designed to intentionally favor one or more players by not respecting fair play policies

Chip Dumping, in Poker Games
Having the data, it is possible to solve this problem through Machine Learning
Using a particular class of neural networks called GNNs

32 of 37

CHIP DUMPING

SOLVE THROUGH GNN

Suppose we have historical data available
Data was generated by a tool called BScreened
It is possible to set static rules for the detection
The rules generate false positives
Manual review is required

Can we optimize with neural networks?

33 of 37