MODULE 4: Chapter 1

Current Trends in Artificial Intelligence

 Artificial Intelligence (AI) is a rapidly growing field in computer science that focuses on creating
machines capable of thinking and acting like humans. AI systems can learn from data, recognize
patterns, make decisions, and solve problems with little or no human involvement.
 Recent advancements in AI are mainly due to faster computers, the availability of large amounts
of data, and improved algorithms. These developments have made AI more accurate, efficient,
and widely used in real-life applications.
 Today, AI plays an important role in engineering and technology. It is used in areas such as
automation, robotics, healthcare, transportation, smart devices, and manufacturing. AI helps
engineers design better systems, reduce errors, and improve performance.
 As AI continues to grow, understanding its current trends is essential for engineering students, as
it opens new opportunities for innovation and future careers.
8.1 AI and Ethical Concerns
 Ethics refers to the study of moral values, duties, and responsibilities that guide human behavior.
It helps us understand what is right and what is wrong.
 Similarly, AI ethics focuses on the moral issues related to the design, development, and use of
artificial intelligence systems such as robots and intelligent machines. AI ethics ensures that AI
technologies are used in a responsible, fair, and safe manner.
 AI ethics can be broadly divided into two main categories:
[Link]
Roboethics deals with the ethical responsibilities of humans while designing, building, using, and
interacting with intelligent machines and robots. It focuses on questions such as:
 How should humans treat robots?
 How should robots be designed to ensure safety?
 How can misuse of robots be avoided?
2. Machine Ethics
Machine ethics focuses on the moral behavior of intelligent machines themselves. It involves developing
Artificial Moral Agents (AMAs) that can make ethical decisions. These agents play an important role
in achieving Artificial General Intelligence (AGI) while following existing legal, social, and moral
rules.
 Responsible: AI should protect human rights and keep personal data safe.
 Accountable: AI systems should be monitored and improved using feedback.
 Transparent: AI should clearly explain how it makes decisions.
 Empowering: AI should help people, businesses, and society grow and improve.
 Inclusive: AI should respect and benefit everyone, not just a few people.
Major Ethical Concerns in Artificial Intelligence
The increasing use of AI systems has raised several important ethical concerns, especially when AI
systems start behaving like humans.
1. Human-like Output vs Human-like Decisions
If AI systems can generate human-like outputs, an important question arises: can they also make
decisions similar to humans?
2. Nature of AI Decisions
Even if AI makes decisions similar to humans, it is important to ask whether these decisions
truly reflect human reasoning, emotions, and values.
3. AI in Financial Decisions
When AI systems decide to approve or reject bank loans, concerns arise about whether such
decisions are fair, transparent, and justified.
4. AI in Education
If AI systems decide whether a student should be admitted to a college, it raises questions about
fairness, equality, and bias in decision-making.
5. Trustworthiness of AI Decisions
Even if AI makes human-like decisions, it is important to consider whether these decisions can
be trusted in critical situations.
AI systems are mainly built using data, mathematical models, and training algorithms. If the data
used for training is incorrect, incomplete, or biased, the AI system may produce unfair or incorrect
decisions.
Such ethical issues are numerous and complex. Therefore, understanding these concerns is essential for
the responsible and successful development of AI technology.
8.1.1 Ethical Use of Artificial Intelligence
 AI systems provide many benefits to businesses, but their use also creates ethical concerns.
 An ethical AI system should be fair and unbiased. This is very important when AI uses
complex methods like deep learning and Generative Adversarial Networks (GANs), which
are often difficult to explain.
 In industries with strict rules, such as banking and finance, AI decisions must be explained
clearly. For example, banks must explain why a loan was approved or rejected.
 However, many AI systems make decisions by analyzing thousands of data factors, which makes
it hard to understand how the final decision was made. When an AI system cannot explain its
decision-making process, it is called Black Box AI.
 Because of this, ensuring transparency and fairness in AI is essential for its ethical use.

 A responsible AI system is explainable, monitorable, reproducible, se cure, human-centered,

unbiased and justifiable (refer Fig. 8.2).
 Explainable: AI should clearly explain how and why it makes decisions.
 Monitorable: AI systems should be regularly checked and controlled by humans.
 Reproducible: AI results should be consistent when the same input is used.
 Secure: AI systems should protect data and prevent misuse or attacks.
 Human-centered: AI should focus on helping humans and respecting human values.
 Unbiased: AI should treat everyone fairly without discrimination.
 Justifiable: AI decisions should be reasonable and ethically acceptable.
Together, these features help build Responsible and Ethical AI systems.
 Formulating laws to regulate AI is very difficult because AI uses many different technologies. If
laws are too strict, they can slow down the growth and development of AI.
 AI is changing very fast, and new advancements can quickly make existing laws outdated. For
example, current privacy laws for phone calls and recorded conversations do not fully address
the challenges created by voice assistants like Amazon Alexa and Apple Siri.
  These digital assistants collect user data mainly to improve their learning and performance, not
to misuse it. However, this data collection can be misused by criminals with harmful intentions.
 Therefore, creating effective and flexible laws for AI is important to ensure safety without
stopping innovation.
8.1.2 Is AI Dangerous? Will Robots Take Over the World?
The rapid growth and increasing capabilities of AI systems have raised concerns about the risks
involved in using AI. Many people fear that AI may replace human jobs or take over human tasks.
However, most researchers believe that super-intelligent AI is unlikely to have human emotions, so it
is not naturally evil or malevolent.
When Can AI Become Dangerous?
AI can become risky mainly in two situations:
1. AI Programmed for Harm
AI systems can be dangerous if they are intentionally programmed to cause destruction.
 Autonomous weapons are examples of AI systems designed to kill.
 If such weapons fall into the hands of people with bad intentions, they could lead to wars, mass
destruction, and loss of human lives.
 These weapons are very difficult to switch off, and once activated, humans may lose control over
them.
 Even narrow AI can be dangerous when given too much autonomy.
2. Misuse or Poorly Defined Goals
AI can also become harmful if its goals are not clearly defined.
 For example, instructing a self-driving car to reach a destination ―as fast as possible‖ without
safety rules may cause accidents.
 AI may complete its task successfully but create serious damage in the process.
 Therefore, AI systems must be continuously monitored and controlled.
Examples of AI Risks
 Autonomous drones with facial recognition and weapons can be misused.
 Self-driving cars connected to the internet can be hacked and cause accidents.
 Hospital equipment connected to the internet can be hacked, putting patients’ lives at risk.
Impact on Employment
 AI and Robotic Process Automation (RPA) may replace many low-skilled jobs.
 It is estimated that around 30% of low-skilled service and BPO jobs may be lost due to
automation.
Legal and Responsibility Issues
Currently, there is no clear legal framework to assign responsibility for AI actions. This raises several
questions:
1. Should AI systems be judged like humans?
2. Who is responsible if an AI system becomes self-learning and autonomous?
 3. Is any error margin acceptable if AI mistakes result in loss of life?
Apart from these two serious concerns, AI also poses some additional threats/risks that calls for
attention. These threats are discussed below

[Link] Risk of Job Automation:

 One of the biggest concerns with Artificial Intelligence is job automation. As AI-powered
robots become smarter and more efficient, they are increasingly replacing humans in many jobs.
For example, some restaurants now use robots as waiters, reducing the need for human
workers.
 This situation creates a serious problem for less-educated and low-skilled workers, as
machines may take over their jobs, leaving them unemployed.
 AI-based automation mainly affects jobs that are predictable, repetitive, and routine in nature.
According to a 2019 Brookings Institution study, around 36 million people are working in
jobs that could soon be automated, as nearly 70% of their tasks can be performed by AI. These
jobs include retail sales, market analysis, hospitality, and warehouse work.
 Recent studies also suggest that white-collar jobs may be heavily affected by automation. A
2018 McKinsey & Company report states that the African American workforce is likely to
be more severely impacted by job automation.
[Link] Algorithms
 Computers work on the GIGO principle (Garbage In, Garbage Out), which means the quality
of output depends on the quality of input data.
 In AI systems, if the training data contains biased or incorrect information, the AI will
produce biased results. This is a serious limitation of AI applications.
 When such biased outputs are used in real-world situations, they can create bigger problems such
as unfair decisions, discrimination, or public distrust.
 For example, a dynamic advertising billboard in Utrecht was shut down because it used
spying software, which caused public outrage due to privacy and bias concerns.
Therefore, it is very important to use clean, fair, and unbiased data while designing AI systems.
3. Too Little Privacy
 When we use AI and IoT systems, a huge amount of data is generated every day. Around 2.5
quintillion bytes of data are added daily, and about 90% of all digital data has been created in
just the last two years.
 Smart systems need large amounts of data to work properly. Because of this, a lot of personal
information about users is collected, which reduces individual privacy.

Once data is collected, users often do not know:

 What data is being collected
 Who is using the data
 For what purpose the data is used
Modern cameras can be combined with facial recognition software to capture details such as age,
gender, ethnicity, facial expressions, and even mood.
For example, in China, some police officers use glasses with facial recognition technology connected to
databases containing images of thousands of suspects, identified based on certain behaviors.
Therefore, the growing use of AI raises serious concerns about privacy and surveillance.
[Link] Becomes Unreliable
 With the growth of AI, fake news and filter bubbles have become common. AI systems are
powerful enough to create realistic faces, write convincing text, generate tweets, edit images,
clone human voices, and run targeted advertisements.
 Because of this, it becomes difficult to identify what is real and what is fake. People may be
misled by false information, which can affect opinions, decisions, and trust in digital content. As
a result, information on the internet can no longer be fully trusted.
 AI systems can be used to change videos, such as turning day scenes into night scenes or
creating realistic faces of people who do not even exist.
  Using tools like Deepfake software, faces can be easily placed onto moving videos. Because of
this, celebrities are often targeted, as fake and inappropriate videos are created using their
faces. Even ordinary people can be blackmailed using edited photos or videos. This type of
crime is called Faceswap video blackmailing. (Refer Fig 8.4)
 AI-generated fake videos are also used in politics to spread false information and influence
public opinion.
 For example, Cambridge Analytica was found to have used data from 87 million Facebook
users to influence voters during President Trump’s election campaign.

8.1.3 Towards Ethical and Trustworthy AI

1. Global Regulation Snapshot

European Union – EU AI Act

 First comprehensive AI law using a risk-based approach.

 Categories:
o Unacceptable risk: banned (e.g., social scoring).
o High risk: strict requirements (e.g., hiring, credit scoring, medical devices).
o Limited risk: transparency duties (e.g., chatbots must identify themselves as AI).
o Minimal risk: largely free (e.g., AI in games).
 High‑risk systems must have:
o Risk management, data quality processes, documentation, logs, transparency to users,
human oversight, and robustness measures.[1]
 Penalties: Up to 35 million EUR or 7% of global turnover for worst violations.
United States

 Executive Order (2023) emphasizes safety testing of large models and the creation of an AI
Safety Institute; NIST releases an AI Risk Management Framework (voluntary but
influential).
Other Countries

 Canada: AIDA focuses on ―high‑impact AI‖ with risk assessments and human oversight.
 UK: Sector‑specific guidance instead of one big AI law.
 China: Rules on algorithmic transparency and security, with emphasis on stability and
national security.
 Singapore: Model AI Governance Framework (voluntary, very practice‑oriented).
2. Organizational Tools for Trustworthy AI
1. AI Principles
o Example: Google’s AI Principles – socially beneficial; avoid unfair bias; safety;
accountability; privacy; scientific excellence; restricted to appropriate uses.
2. AI Impact Assessments & Risk Assessments
o Similar to environmental impact assessments, but for AI.
o Ask: Who can be harmed? How severe? What safeguards exist?
3. Red‑Team Exercises & Algorithmic Audits
o ―Attack‖ the model with adversarial inputs or bias tests to see how it fails.
4. Privacy‑Preserving Techniques
 o Differential privacy, federated learning, homomorphic encryption, secure multi‑party
computation.

3. Industry Examples
 Healthcare: Hospital deploying an AI triage system must log decisions, explain rationales to
clinicians, and allow doctors to override.[1]
 Finance: A bank’s credit‑scoring AI is regularly audited to compare approval rates and
default rates across gender and demographic groups.[1]
 Technology Platforms: Social media companies implement internal ethic boards and
external audits for recommender systems to reduce harmful content amplification.[1]
Think: If you were designing an AI loan system for a public sector bank, what simple
fairness checks would you build into the pipeline?

8.1.4 Why is Ethical AI Important?

1. High-Stakes Domains
1. Healthcare
o Diagnostic AI influences life‑or‑death decisions and must protect patient data and
ensure equity across populations.[1]
o Example: IBM Watson for Oncology raised concerns when its recommendations were
not always aligned with oncologists’ judgments, showing the need for rigorous
validation.[1]
2. Criminal Justice

o COMPAS bias case: higher false‑positive risk scores for Black defendants.[1]
o Facial recognition misidentification led to wrongful arrest of Robert Williams in
Detroit (2020).[1]
3. Employment & Recruitment
o Amazon’s hiring AI learned from male‑dominated historical data and downgraded CVs
mentioning ―women’s‖, leading to scrapping of the system.[1]
4. Finance & Education
o Credit scoring can lock people out of loans; AI grading can influence students’
futures.[1]

2. Consequences of Unethical AI
1. Erosion of Privacy
o Large‑scale data collection enables surveillance capitalism and behavior
manipulation.[1]
2. Unexplainable Decisions
o When someone is denied a loan, job, or bail with no understandable rationale, trust in
institutions erodes.[1]
3. Lack of Accountability
o ―The algorithm did it‖ becomes a convenient excuse; victims struggle to get
compensation or correction.[1]
4. Real-World Harm Examples

o Health AI in US hospitals showing racial bias in who receives extra care.[1]

o Wrongful arrests and harassment due to facial recognition errors.[1]
 Think: Should AI systems in public services (policing, welfare, education) be more
transparent than commercial AI? Why?

8.2 AI as a Service (AIaaS)

AI as a Service (AIaaS) lets companies use ready-made AI tools instead of building their own
expensive AI systems.
Just like software can be provided over the cloud, AI services can also be accessed online. This allows
even small businesses to use affordable AI solutions to improve their performance significantly.
8.2.1 Factors Triggering Growth of AIaaS
AI as a Service (AIaaS) is growing fast, with many startups leading the way. Recent IT advancements
have made AIaaS more popular:
Technological factors:
1. Cloud platforms offer affordable options for managing enterprise data.
2. Data storage is now cheaper, reliable, and improved.
3. Streaming devices and IoT generate huge amounts of data that can be analyzed for valuable
insights.
4. Automated data management, analytics, and BI tools have changed how businesses analyze
information.
Business reasons for adopting AIaaS:
1. Building AI infrastructure is expensive.
2. There is a lack of trained programmers.
3. Hiring programmers for AI implementation is very costly.
4. Many companies do not have enough data to analyze on their own.
8.2.2 The Growth of AIaaS
With cloud services becoming widely accessible, AI has grown rapidly. AIaaS allows many companies
to collect, store, and analyze large amounts of data without building their own AI systems.
Benefits of AIaaS for businesses:
1. Focus on their core work instead of becoming AI experts.
2. Reduce costs for operations and infrastructure. AI systems need expensive machines and GPUs,
which AIaaS provides at a lower cost.
3. Lower investment risk.
4. Faster deployment of AI solutions.
5. Pre-built algorithms make complex data management easy, even for small businesses.
6. On-demand services give a competitive advantage.
7. Better insights from analyzing data trends.
8. Cost-effective, flexible, and transparent solutions using templates, pre-built models, and drag-
and-drop tools. Companies don’t need to maintain a full AI team.
9. Strategic flexibility: pay only for what is used, saving costs for small-scale or occasional use.
 10. Scalability: start with small projects and expand services as needed. Companies can scale up or
down depending on demand.
In short, AIaaS makes machine learning accessible, affordable, and flexible for businesses of all
sizes.

Figure 8.7 summarizes the advantages and shortcomings of providing AI as a Service.

8.2.3 Challenges of AIaaS
While AIaaS has many benefits, it also comes with challenges:
1. Reduced security: Companies must share data with third-party vendors, so strong security
measures are needed to prevent theft or misuse.
2. Reliance on vendors: Organizations depend on third parties for services, which can cause
delays in getting results.
3. Reduced transparency: AIaaS often works as a ―black box.‖ Companies get the output but
cannot see how it works, which can create confusion about data stability or results.
4. Data governance issues: Some industries have strict rules about data storage, which may
prevent them from fully using AIaaS on third-party clouds.
8.2.4 Vendors of AIaaS
Many companies provide AI and machine learning services. Organizations can compare these services
and costs to choose the best option for their needs.
Cloud AI providers offer GPU-based processing, maintenance, and hardware support for different
tasks. Major AIaaS vendors include:
 Amazon Web Services (AWS)
 Microsoft Azure
 Google Cloud Platform (GCP)
These vendors provide different bots, APIs, and machine learning frameworks for businesses.
In addition, many startups are offering AIaaS, and larger companies sometimes acquire smaller ones to
expand their services.
Organizations choose an AI service based on:
 Their goals
 Business size
 Budget
 Technical capabilities of their team
 Amount of data to process
Examples of AIaaS Platforms
1. MonkeyLearn
o Simplifies text analysis with no-code tools.
o Users can use pre-trained models (e.g., survey analyzer) or build custom models for
sentiment, keywords, and topic detection.
o Models can be integrated with apps via point-and-click or API.
2. IBM Watson
o Offers AI tools and pre-built applications like Watson Assistant (virtual assistants) and
Watson Natural Language Understanding (advanced text analysis).
o Watson Studio lets users build, train, and deploy machine learning models across any
cloud, even without AI expertise.
3. Microsoft Azure
o Cloud platform providing AI and machine learning tools.
o Azure Cognitive Services adds computer vision or text extraction via APIs.
o Azure Bot Service allows integration of bots into applications.
4. Google Cloud ML
o Supports data scientists and developers to create and deploy ML projects with big data.
o AutoML allows training custom models for text, image classification, translation, etc.
o Provides tools to visualize data, test scenarios, and evaluate model performance.
AI Platform-as-a-Service (AIPaaS)
AIPaaS is a complete cloud-based solution that lets businesses use AI services on a pay-per-use basis.
Companies can access third-party APIs to manage and analyze data intelligently without building their
own AI systems.
The market size shows its growth potential:
 SaaS (Software-as-a-Service): valued at $133 billion.
 Advanced Platform-as-a-Service: expected to reach $11 billion by 2023 and $88.5 billion by
2025.
8.3 Recent Trends in AI
 This section covers the latest trends in AI. These are areas where extensive research is
currently being done.
8.3.1 Collaborative Systems
Research is being done to create collaborative systems, where humans and AI work together to
overcome each other’s limitations and enhance human abilities. Some examples include:
1. Foldit game – Humans and AI fold simulated proteins to understand diseases. AI handles tasks
it is best at, while humans use intuition and creativity.
2. Chess collaboration – Two amateur players and an AI system beat grandmasters and
supercomputers.
3. Business – In 2014, a Japanese venture capital company, Knowledge Ventures, included an AI
system on its board of directors.
4. Military drones – UAV drones take front-line actions while humans provide virtual support.
5. Talos exoskeleton suit – Augments soldiers with power assistance, vital monitoring, and
HUDs, similar to Ironman-style collaboration.
Debate:
 Many believe humans and machines can be effective partners.
 Others fear over-reliance on AI could cause job loss and make people less skilled or lazy.
Research involving 1,500 companies shows that when AI is used to replace employees, it often brings
only short-term gains.
Long-term benefits happen when humans and machines work together, combining their strengths:
 Humans: leadership, teamwork, creativity, social skills
 Machines: speed, scalability, quantitative analysis
AI can enhance our decision-making, analytical abilities, and creativity when used alongside
humans.
Humans Assisting Machines
In human-machine collaboration, humans have three key roles:
1. Train machines to perform tasks.
2. Explain outcomes, especially when results are unexpected or controversial.
3. Ensure responsible use so machines do not harm people.
 Example: The EU’s GDPR gives consumers the right to an explanation for algorithm-based
decisions (like loan approvals). This not only promotes responsible AI but also creates about
75,000 jobs to manage GDPR compliance.
Humans also need to monitor AI systems to ensure they are safe and working correctly:
 Self-driving cars must avoid hitting objects or people.
 Companies like Apple use AI to improve user experience, but employees must make sure
privacy is protected and data is used responsibly.
8.3.2 Machines Assisting Humans
Smart machines help humans in three ways: amplify, interact, and embody.
 Amplify: AI boosts creativity, analysis, and decision-making by providing the right information
at the right time.
 Interact: AI communicates effectively with employees and customers.
  Embody: AI assists humans in daily tasks, even at home, using digital assistants.

Example: SEB, a Swedish bank, uses Aida, a virtual assistant, to:

 Help customers open accounts and make cross-border payments.
 Interpret customer emotions (satisfied or frustrated) to improve service.
 Direct unresolved issues (30% of cases) to human representatives and learn from these
interactions to solve similar problems in the future.
Case study Question: Collect information about how Autodesk’s Dreamcatcher AI enhances the
imagination of even exceptional designers.
Embodying
AI systems can be embodied in machines, like robots, to enhance human abilities.
 Using sensors, motors, and actuators, these machines can recognize people and objects and
work safely in factories, warehouses, and labs.
 Cobots (collaborative robots) handle repetitive, heavy, or dangerous tasks, allowing humans
to focus on tasks requiring skill and dexterity.
Examples:
 Hyundai uses exoskeletons—wearable robotic devices that adapt to workers in real time, giving
them superhuman strength and endurance.
 Mercedes-Benz uses cobot arms to customize cars based on real-time customer choices at
showrooms.
8.3.3 Algorithmic Game Theory and Computational Social Choice
 With the rise of Internet connectivity and faster online speeds—trends that were especially
accelerated during the pandemic—more people are playing strategic games like chess, checkers,
poker, and solitaire. These games operate under a clear set of rules, which means that machines
designed to play them must also follow these rules in a predictable and ―ethical‖ manner.
 This requirement led to the development of game theory, a field that studies strategic
interactions between rational agents. In multi-agent situations (i.e., scenarios involving more
than one participant), game theory helps in choosing an optimal action from a set of possible
options, while considering that:
 Our choice will affect the opponent’s decisions.
 The opponent’s decisions will, in turn, affect our own outcomes.
Thus, game theory provides a framework for reasoning about interdependent decision-making, which is
crucial in both human and machine-based gameplay.
Von Neumann is credited with the invention of game theory. The concept of game theory can be
understood by studying the following five categories of games:
1. Cooperative vs Non-cooperative Games:
o In cooperative games, participants can form alliances to increase their chances of
winning.
 o In non-cooperative games, no alliances are allowed, and each participant acts
independently.
2. Symmetric vs Asymmetric Games:
o In symmetric games, all participants share the same goals, though their strategies to
achieve them may differ.
o In asymmetric games, participants have different, and sometimes conflicting, goals.
3. Perfect vs Imperfect Information Games:
o In perfect information games (like chess), all players can observe the moves of other
players.
o In imperfect information games (like card games), some moves or information are
hidden from the players.
4. Simultaneous vs Sequential Games:
o In simultaneous games, multiple players act at the same time without knowledge of
others’ choices.
o In sequential games (like most board games), players make moves one after another,
with each aware of previous actions.
5. Zero-sum vs Non-zero-sum Games:
o In zero-sum games, one player’s gain is exactly another player’s loss.
o In non-zero-sum games, multiple players can benefit simultaneously, even when another
player gains.
When applying AI techniques in game theory, two concepts are particularly important:
 Nash Equilibrium: A state where no player can improve their outcome by changing their
strategy alone.
 Inverse Game Theory: A method to deduce players’ objectives or preferences based on
observed behavior in a game.
Nash Equilibrium
The Nash equilibrium is a condition in a game where all players recognize that no alternative strategy
would improve their outcome, given the strategies chosen by others. In other words, no player can gain
by unilaterally changing their current strategy. Any new decision by a player would not result in a
better outcome.
Example: Prisoner’s Dilemma (Figure 8.21)
Consider the classic Prisoner’s Dilemma:
 Two criminals are arrested and kept in separate confinement, unable to communicate with each
other.
 The possible outcomes are:
1. If one prisoner confesses and the other remains silent:
 The confessor goes free.
 The silent prisoner is sentenced to 10 years in prison.
 2. If both prisoners remain silent:
 Each receives 1 year in prison.
3. If both confess:
 Each receives 5 years in prison.
In this scenario, the Nash equilibrium occurs when both prisoners betray each other. Although
mutual silence would result in a shorter combined sentence, neither prisoner can improve their personal
outcome by changing their own strategy alone.

Inverse Game Theory

While game theory focuses on analyzing a game to optimize the possible outcomes for its players,
inverse game theory takes the opposite approach: it designs a game based on observed players’
strategies and goals. This is particularly important for creating environments for AI agents, where
understanding how agents act allows designers to structure interactions effectively.
Practical Example
A common application of inverse game theory is in Generative Adversarial Networks (GANs):
 A GAN consists of two neural networks that compete with each other:
1. The generator tries to create data that mimics real data.
2. The discriminator evaluates the data and determines if it is real or generated.
 Through this competitive process, both networks improve over time.
 Nash equilibrium is generally easier to achieve in symmetric games, where players have
similar strategies and goals.
 In contrast, asymmetric games—where players have different or conflicting goals—are more
common in real-world scenarios, making equilibrium analysis more complex.
  Generative Models: These models take input features and attempt to learn the underlying
distribution of the data. By understanding how the data is generated, they can create new
samples that resemble the original data.
o Example in GANs: The generator uses input features to produce new data samples that
mimic the main characteristics of the original dataset.
 Discriminative Models: These models focus on predicting the class or label of a given input
sample. They do not attempt to model the data distribution, but instead learn the boundaries
between classes.
 In a GAN, the generative model creates realistic samples, while the discriminative model
evaluates them, guiding the generator to improve its output. This interplay helps the system learn
effectively.
Role of the Discriminative Model in GANs
 In a Generative Adversarial Network (GAN), both the original samples and the newly
generated samples are passed to the discriminative model. The role of this model is to identify
which samples are genuine (real) and which are fake.
A common application of GANs is in image generation:
 The generator creates new images (as given in Fig. 8.23) that resemble real images.
 The discriminator evaluates these images to determine whether they are real or fake.
 Through this adversarial process, both models improve over time, resulting in highly realistic
generated images.
GANs in Game Playing
GANs are also extensively used in game-playing scenarios, where the two models act as competing
players:
 One player (the generator) creates fake samples to challenge or confuse the other player.
 The other player (the discriminator) tries to correctly identify which samples are real and
which are fake.
The game is played iteratively, and in each iteration:
 The models update their learning parameters to minimize their respective losses.
 This process continues until the system reaches a Nash equilibrium, where neither player can
improve their strategy without the other changing theirs.
This iterative adversarial process allows both models to continuously improve, resulting in highly
refined strategies or generated outputs.
8.3.4 Multi-Agent Reinforcement Learning (MARL)
Reinforcement Learning (RL) is a framework in which an agent learns through interaction with its
environment. Traditionally, RL is based on Markov Decision Processes (MDPs), where:
 An agent is placed in a stochastic, stationary environment.
 It learns an optimal policy through a reward/punishment mechanism.
 Over time, the agent converges to a satisfactory policy that maximizes cumulative rewards.
Example: Traffic Flow Optimization with MARL and Game Theory
Consider the scenario of AI-powered self-driving cars in a city:
 Single-agent case: Each car interacts with the external environment to optimize its own path,
improving traffic flow individually.
 Multi-agent case: When multiple cars act as a group, conflicts can arise. For example, two cars
may both choose the same optimal route, creating congestion.
This situation can be effectively modeled using game theory:
 Each car acts as a player in the game.
 The strategies correspond to the routes that each car may choose.
 The Nash equilibrium represents a state where all cars have chosen routes in such a way that no
single car can improve its travel time by changing its route unilaterally.
By applying MARL and game theory, self-driving cars can learn to collaborate implicitly, reducing
conflicts and optimizing traffic flow across the city.
Challenges in Multi-Agent Environments
When multiple agents are placed in the same environment, the standard RL assumptions no longer hold
because:
 The learning process now depends not only on interactions between an agent and the
environment but also on interactions among the agents themselves.
 The number of possible interactions grows exponentially as the number of agents increases,
making modeling and learning much more complex.
Approaches to Simplify MARL
To handle this complexity, Multi-Agent Reinforcement Learning (MARL) models often use
techniques like Mean Field Scenarios (MFS):
 MFS reduces complexity by assuming a priori that all agents have similar reward functions.
 This allows the system to approximate the effects of all other agents on a given agent through a
mean field, rather than modeling every individual interaction.
By using such approximations, MARL becomes computationally feasible while still capturing the
essential dynamics of multi-agent interactions.