Artificial Intelligence
Assignment: 02
Submitted by: Sakina Batool
Sap Id: 70132094
Question 1
Comparison of Intelligent Agents
Intelligent agents are entities that perceive their environment through sensors
and act upon it through actuators. Here is how they differ based on their
internal complexity:
1. Simple Reflex Agents
• Working Principle: These agents follow the Condition-Action rule (\bm{if
\text{ state } X \text{ then action } Y}). They ignore the history of the
environment and respond only to the current percept.
• Contrast: They are the least "intelligent" because they fail if the
environment is even partially unobservable.
2. Model-Based Reflex Agents
• Working Principle: These maintain an internal state (model) that tracks
parts of the world they cannot see right now. They use "knowledge about
how the world evolves" to handle partial observability.
• Contrast: Unlike simple reflex agents, they have a "memory" of the
environment's state over time.
3. Goal-Based Agents
• Working Principle: These agents act to achieve a specific target state. They
use searching and planning algorithms to find a path to their goal.
• Contrast: While reflex agents act on rules, goal-based agents act on
outcomes. They are more flexible because the goal can be changed without
rewriting all the rules.
4. Utility-Based Agents
• Working Principle: These use a Utility Function to measure "happiness" or
"success." They don't just want to reach a goal; they want to reach it in the
best (fastest, cheapest, or safest) way.
• Contrast: They distinguish between different paths to a goal based on a
preference scale.
5. Learning Agents
• Working Principle: These consist of four components: a learning element
(to make improvements), a performance element (to act), a critic (to give
feedback), and a problem generator (to suggest new experiences).
• Contrast: They can operate in initially unknown environments and become
more competent over time.
Question 2
Major Branches of Artificial Intelligence
1. Machine Learning (ML)
• Real-World Application: Credit Card Fraud Detection.
• AI Principles: This demonstrates the principle of Learning and Pattern
Recognition. Instead of hard-coding every possible fraud scenario, the
system analyzes millions of transactions to recognize statistical anomalies
that represent "theft."
2. Natural Language Processing (NLP)
• Real-World Application: Real-time Language Translation (e.g., Google
Translate).
• AI Principles: This demonstrates Perception and Communication. It
involves understanding the syntax, semantics, and context of human
language to map meaning from one symbolic representation (English) to
another (Urdu).
3. Computer Vision (CV)
• Real-World Application: Facial Recognition Systems (e.g., FaceID on
iPhones).
• AI Principles: This demonstrates the principle of Visual Perception. It
involves feature extraction (identifying eyes, nose, and distance between
them) and classification to identify a specific entity from a visual input.
Question 3
ChatGPT vs. Gemini (formerly Bard) and Technical Insights
Functionality, Performance, and Scalability
• Functionality: ChatGPT (OpenAI) excels in creative writing, coding, and
complex reasoning. Gemini (Google) is more integrated with real-time web
searching and the Google ecosystem (Docs, Gmail).
• Performance: ChatGPT often scores higher in logic-heavy benchmarks
(GPT-4o), while Gemini is noted for its multimodal speed and ability to
process very large "context windows" (up to 2 million tokens).
• Scalability: Both use massive cloud infrastructures (Azure for OpenAI,
Google Cloud for Gemini) allowing them to serve millions of users
simultaneously via API.
Is it a State Space Search Problem?
No, it is primarily a Probabilistic Sequence Prediction problem.
• Justification: In a State Space Search (like Chess), there is a clearly
defined start state, a goal state, and specific rules to transition between
them. You search for the "best path."
• LLM Reality: ChatGPT does not "search" for an answer in a pre-defined
space. It generates tokens one by one based on probabilities learned during
training. It uses a Transformer architecture to predict the most likely "next
word" given the context.
Technical Implementation from an AI Perspective
ChatGPT is built on the Transformer Architecture using a process called
RLHF (Reinforcement Learning from Human Feedback):
1. Pre-training: The model learns grammar and facts from a massive
dataset (Petabytes of text).
2. SFT (Supervised Fine-Tuning): Human trainers provide "ideal"
conversations to teach the model how to follow instructions.
3. RLHF: The model generates multiple answers, and humans rank them. A
Reward Model is trained on these rankings to guide the final model toward
helpful and safe responses.
Testing ChatGPT: AI & NLP (Urdu Example)
If you ask a question in Urdu, such as "( "مصنوعی ذہانت کیا ہے؟What is AI?),
ChatGPT uses its cross-lingual embeddings. It doesn't just translate; it
understands the "token" representations in Urdu and generates a response
based on the concepts it learned in its multi-lingual training set.