Abdulhakim Bashir
My E-Portfolio based on work carried out on my Msc Program on Artificial Intelligence and Machine Learning at the University of Essex.
Intelligent Agents 2025
Collaborative Discussion 1: Agent Based Systems
Discussion on fundamental concepts of agent-based systems and their applications
Unit 1: Introduction to Agent-Based Computing
This unit introduces the fundamental concepts of intelligent agents, their environment, and basic agent architectures.
1.1 Key Concepts
Intelligent agents have four main characteristics:
- Autonomy: Operating independently without direct human control
- Social ability: Interacting with other agents and humans
- Reactivity: Perceiving and responding to changes in the environment
- Proactivity: Taking initiative to achieve goals, not just reacting
1.2 Learning Outcomes
Key outcomes from this unit include:
- Understanding the definition of intelligent agents and their properties
- Identifying different types of agent environments
- Distinguishing between reactive, deliberative, and hybrid architectures
- Applying basic agent design principles to practical problems
1.3 Reflection Notes
The agent paradigm provides a useful framework for AI systems. There are important connections between human cognition and agent architecture that inform design approaches (Wooldridge, 2009). This is particularly evident in how agents can be designed to implement decision-making processes.
A significant challenge in agent design is balancing reactivity with goal-directed behavior (Russell & Norvig, 2020). Agents that are too reactive may not achieve long-term goals, while those too focused on goals might miss important environmental changes. Finding the right balance is critical for effective application (Wooldridge, 2009). Future research could explore how agents might adjust this balance based on their current situation.
References
- Wooldridge, M.J. (2009). An Introduction to Multiagent Systems. Chichester: John Wiley & Sons. Chapter 2.
Unit 2: Introducing First Order Logic
This unit introduces first order (predicate) logic as a mathematical foundation for knowledge representation in intelligent agent systems.
2.1 Key Concepts
- Predicates and Relations: Expressions describing object properties and relationships
- Quantifiers: Universal (∀) and existential (∃) operators for statements about collections
- Inference Rules: Methods for deriving new statements from existing ones
2.2 Learning Outcomes
- Understanding core elements of first order logic
- Creating and reasoning over first order logic
- Recognizing the relationship between first order logic and natural language
2.3 Reflection Notes
First order logic (FOL) offers significant advantages over simpler logical forms. Russell and Norvig (2021) note FOL "can express facts about some or all of the objects in the universe," making it more powerful than propositional logic for complex domains.
FOL's expressiveness comes from predicates representing objects, properties, and relations, while quantifiers enable reasoning about collections. This makes FOL especially valuable for agent systems that interact with humans and need to make generalizations across environments (Russell and Norvig, 2021).
References
- Russell, S. & Norvig, P. (2021) Artificial Intelligence: A Modern Approach. 4th ed. Harlow: Pearson Education. Chapter 8.
Unit 3: Agent Architectures
Agent architectures define how agents perceive, reason, and act. The unit examines different architectural models and their performance across problem domains.
3.1 Key Concepts
- Symbolic Reasoning Agents: Knowledge-based systems using explicit symbolic representations
- Reactive Architectures: Systems generating responses directly from environmental stimuli
- BDI (Belief-Desire-Intention): Agents with mental states representing knowledge, goals, and commitments
- Subsumption Architecture: Layered approach creating complex behaviors from simple components
3.2 Learning Outcomes
- Evaluating different agent architectures
- Selecting appropriate architectures for specific tasks
- Understanding intentions versus desires in agent design
3.3 Reflection Notes
Maes (1991) defines architecture as a methodology that "specifies how the agent can be decomposed into a set of component modules and how these modules should interact" to transform sensor data into actions.
Symbolic agents use explicit representations but face challenges in real-world complexity. Brooks (1991, p.140) identifies the "transduction problem" of translating reality into symbols and performing manipulations efficiently.
Brooks' (1991) subsumption architecture demonstrates how complex behaviors emerge from simple components without internal models, proving effective for real-time applications.
The BDI model structures reasoning around beliefs, desires, and intentions. Bratman et al. (1988) describe intentions as "conduct-controlling pro-attitudes" that agents maintain without constant reconsideration, enabling goal persistence with environmental responsiveness.
References
- Maes, P. (1991) The Agent Network Architecture (ANA). SIGART Bulletin 2(4): 115-120.
- Kaelbling L. P. (1991) A Situated-automata Approach to the Design of Embedded Agents. ACM SIGART Bulletin 2(4): 85-8.
- Bratman, M.E., Israel, D.J. & Pollack, M.E. (1988) Plans and Resource‐bounded Practical Reasoning. Computational Intelligence 4(3): 349-355.
- Brooks, R.A. (1991) Intelligence Without Representation. Artificial Intelligence 47(1-3): 139-159.
- Reynolds, C.W. (1987) Flocks, Herds, and Schools: A Distributed Behavioral Model. Computer Graphics 21(4): 25-34.
Unit 4: Hybrid Agent Architectures
Hybrid agent architectures combine reactive and deliberative components to balance responsiveness with reasoning.
4.1 Key Concepts
- Hybrid Architectures: Systems integrating reactive and deliberative components
- Layering: Horizontal (parallel) or vertical (sequential) information flow in hierarchical structures
- Examples: TouringMachines (three-layer) and InteRRaP (vertical structure)
4.2 Learning Outcomes
- Critically assessing agent architectures for specific problems
4.3 Reflection Notes
Hybrid architectures combine responsive behavior with goal-oriented reasoning, addressing limitations of both reactive and deliberative approaches. Most use layered structures with different patterns for information flow and control (Wooldridge, 2009).
References
- Wooldridge, M.J. (2009) An Introduction to Multiagent Systems. 2nd ed. Chichester: John Wiley & Sons. Chapter 5.
Unit 5: Agent Communication
Explores how agents communicate with each other, focusing on speech act theory, agent communication languages, and ontologies.
5.1 Key Concepts
- Speech Acts: Utterances that perform actions beyond conveying information
- KQML: Knowledge Query and Manipulation Language for agent communication
- Ontologies: Formal knowledge representations supporting shared understanding
- CID: Correspondence Inclusion Dialogue for ontology alignment
5.2 Learning Outcomes
- Developing and working with ontologies
- Creating inter-agent communications using relevant languages
5.3 Reflection Notes
Speech act theory treats language as action. Searle (1969, p.16) states that "speaking a language is engaging in a rule-governed form of behavior," emphasizing that communication changes the world state, not just transfers information.
KQML implements speech act theory by separating performatives (actions) from propositional content, reflecting Searle's distinction between illocutionary force and content (1969). This enables agents to interpret both message content and intended actions.
Ontologies ensure agents share common understanding of terms. Without alignment, agents may use identical terms differently or different terms for the same concepts. As Payne and Tamma (2014) explain, "semantic heterogeneity can impede meaningful communication" between agents with different ontological assumptions.
References
- Searle, J.R. (1969) Speech Acts: An Essay in the Philosophy of Language. Cambridge: Cambridge University Press. Pages 1-53.
- Payne, T.R. & Tamma, V. (2014) 'Negotiating over ontological correspondences with asymmetric and incomplete knowledge', AAMAS, 13(1), pp. 517-524.
Collaborative Discussion 2: Agent Communication Languages
Discussion on agent communication languages and protocols
Unit 6: Working Together
Focuses on practical applications of agent communication languages, specifically KQML and KIF.
6.1 Key Concepts
- KQML: Knowledge Query and Manipulation Language - the protocol layer
- KIF: Knowledge Interchange Format - the content layer
- Performatives: Speech act types defining message intentions
- Agent Dialogues: Structured conversations between agents
6.2 Learning Outcomes
- Developing dialogues using KQML/KIF
- Understanding ontologies for knowledge sharing
- Evaluating approaches to agent communication
6.3 Reflection Notes
KQML provides a standardized framework implementing speech act theory, enabling diverse agents to interact regardless of internal architecture (Finin et al., 1994). The separation into content, message, and communication layers mirrors speech act theory's distinction between illocutionary force and propositional content. Implementation challenges include semantic ambiguity, formal semantics issues, and practical concerns like message routing and security.
References
- Finin, T., Fritzson, R., McKay, D. & McEntire, R. (1994) 'KQML as an agent communication language', CIKM, 1, pp. 456-463.
Unit 7: Natural Language Processing (NLP)
This unit examines NLP technologies for intelligent agents.
7.1 Key Concepts
- Linguistic Analysis: Syntax, semantics, pragmatics
- Statistical Approaches: Word embeddings, vector models
- Parsing: Formal grammars, constituency and dependency methods
7.2 Learning Outcomes
- Evaluating NLP system development challenges
- Understanding core NLP principles
7.3 Reflection Notes
NLP challenges stem from language complexity. Mikolov et al. (2013, p.3111) discuss "distributed representations" for handling ambiguity. Word2Vec enables semantic operations like "king - man + woman ≈ queen" (Mikolov et al., 2013, p.3112).
Hearst (1992, p.539) pioneered pattern-based relationship extraction, while dependency parsing captures "interlinking dependencies of words" (Aqab and Tariq, 2020, p.140). Despite advances, systems struggle with implicit meaning and cultural references.
References
- Mikolov, T., et al. (2013) Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 1, pp. 3111-3119.
- Hearst, M.A. (1992) Automatic acquisition of hyponyms from large text corpora. COLING, 2, pp. 539-545.
- Aqab, S. & Tariq, M.U. (2020) Handwriting recognition using artificial intelligence neural network and image processing. IJACSA, 11(7), pp. 137-146.
Unit 8: Understanding Natural Language Processing (NLP)
This unit provides practical exploration of NLP technologies and techniques.
8.1 Key Concepts
- Word2Vec Models: Neural network techniques for word embeddings
- Vector Space Semantics: Word representation in multi-dimensional space
- Constituency Parsing: Hierarchical analysis of sentence structure
8.2 Learning Outcomes
- Explaining key elements of NLP models
- Engaging with worked NLP examples
- Developing parse trees for language understanding
8.4 Reflection Notes
Word2Vec transforms words into vectors where proximity indicates semantic similarity, capturing relationships that symbolic approaches cannot represent. This data-driven approach uses neural networks to learn from context without explicit linguistic rules (Zimmerman, 2019).
Constituency parsing complements vector-based semantics by revealing grammatical structure through hierarchical decomposition. Parse trees resolve ambiguities impossible to address at word level alone, as in "the man saw the dog with the telescope," which has multiple interpretations depending on prepositional phrase attachment (Zimmerman, 2019).
References
- Zimmerman, V. (2019) 'Getting to Grips with Parse Trees', Towards Data Science.
Unit 9: Introduction to Adaptive Algorithms
This unit introduces adaptive algorithms, with a focus on Artificial Neural Networks and Deep Learning.
9.1 Key Concepts
- Artificial Neural Networks (ANNs): Computational models inspired by biological neural networks
- Deep Learning: Advanced neural networks with multiple hidden layers
- CNNs: Specialized networks for processing grid-like data such as images
- RNNs: Networks with feedback connections for sequential data
9.2 Learning Outcomes
- Understanding core concepts of artificial neural networks
- Appraising the relative strengths of these techniques
- Applying intelligent agent techniques to real-world problems
9.3 Reflection Notes
ANNs represent a paradigm shift from explicitly programmed rules to systems learning directly from data. Deep learning extends this with multiple hidden layers enabling hierarchical representations. Huang (2013) notes deep learning excels at "unsupervised feature learning" without human guidance.
CNNs have revolutionized computer vision with specialized layers mimicking the visual cortex. Microsoft's system demonstrated how CNNs can surpass humans, achieving a 4.94% error rate compared to humans' 5.1% on ImageNet (Thomsen, 2015). RNNs address sequential processing with feedback connections, enabling advances in language tasks.
References
- Huang, X. (2013) 'Andrew N.G on Deep Learning, Self-Taught Learning and Unsupervised Feature Learning'.
- Thomsen, M. (2015) 'Microsoft's Deep Learning Project Outperforms Humans In Image Recognition', Forbes.
Collaborative Discussion 3: Deep Learning
Discussion on deep learning applications in intelligent agents
Unit 10: Deep Learning in Action
This unit explores practical applications of Deep Learning technologies and their societal impacts.
10.1 Key Concepts
- Applied Deep Learning: Real-world neural network implementations
- Transfer Learning: Leveraging pre-trained models for new problems
- Socio-Technical Impact: Broader societal implications
- Ethical AI Deployment: Responsible implementation considerations
10.2 Learning Outcomes
- Understanding technology development and current limitations
- Appraising products based on socio-economic and ethical impacts
- Explaining the importance of data in technology delivery
10.3 Reflection Notes
Deep Learning has rapidly transitioned from research to commercial applications. The World Economic Forum (2022) notes these technologies can "improve productivity and boost business" through cognitive task automation and enhanced decision-making.
Current limitations include massive data requirements and "black box" interpretability challenges. Socio-economic impacts balance efficiency benefits against concerns about job displacement and bias amplification. Data quality remains critical, driving innovations like transfer learning. Ethical deployment requires balancing innovation with potential negative impacts, particularly for vulnerable populations.
References
- World Economic Forum. (2022) 'How Deep Learning can improve productivity and boost business'.
Unit 11: Intelligent Agents in Action
This unit integrates agent-based computing, adaptive algorithms, and deep learning to examine practical applications in various sectors.
11.1 Key Concepts
- Industry 4.0: Smart manufacturing, automation, and data exchange
- Digital Twins: Virtual replicas enabling monitoring and optimization
- Smart Shop Floor: Manufacturing environments with coordinating systems
- Agent-Based Financial Modeling: Multi-agent simulation of markets
11.2 Learning Outcomes
- Applying learned concepts to specific sectors
- Understanding technology-driven efficiency
- Evaluating new approaches in context
11.3 Reflection Notes
Industry 4.0 represents a manufacturing paradigm shift. Foit (2022, p.2) notes agent-based modeling "offers a range of advantages in simulating complex manufacturing systems." This enables flexible production that adapts without central control.
Wang et al. (2016, p.159) propose a "self-organized multi-agent system with big data-based feedback" where machines function as autonomous agents. Digital twins provide virtual replications for optimization, while in finance, the Bank of England (n.d.) explains agent-based models help "understand the economy from the bottom up," simulating emergent phenomena like market crashes.
References
- Bank of England. (n.d.) Agent-based models: understanding the economy from the bottom up.
- Foit, K. (2022) 'Agent-based Modelling of Manufacturing Systems in the Context of "Industry 4.0"', Journal of Physics: Conference Series, 2198, p. 012064.
- Wang, S., et al. (2016) 'Towards smart factory for industry 4.0: a self-organized multi-agent system with big data based feedback and coordination', Computer Networks, 101(4), pp. 158-168.
Unit 12: The Future of Intelligent Agents
This unit explores future directions of intelligent technologies, particularly deep learning and AI systems.
12.1 Key Concepts
- Technological Forecasting: Methodologies for predicting intelligent agent capabilities
- AI Alignment: Ensuring AI systems act according to human values
- Interpretable AI: Making AI decision-making transparent and understandable
- Transformative AI: Systems with potential to alter social, economic, and political structures
12.2 Learning Outcomes
- Appraising current technologies
- Assessing potential evolution based on today's examples
- Evaluating ethical and social considerations
12.3 Reflection Notes
Multimodal learning systems suggest a future with broader contextual understanding. AI alignment represents a critical challenge as systems become more autonomous. Nasim et al. (2022, p.55) note "the gap between AI capabilities and our ability to ensure their safe and beneficial use" continues to widen.
Interpretability challenges grow as neural architectures become more complex. Nasim et al. (2022, p.58) highlight that AI failures often involve systems where "the decision-making process lacks transparency." The evolution of intelligent agents will be shaped by technical advances alongside cultural, economic, and regulatory factors.
References
- Nasim, S.F., Ali, M.R. & Kulsoom, U. (2022) 'Artificial Intelligence Incidents & Ethics A Narrative Review', IJTIM, 2(2), pp. 52-64.