01Understanding Artificial Intelligence
AI research builds computational systems that perform tasks associated with intelligence: solving problems, understanding language, recognizing patterns, planning and making decisions under uncertainty.
A widely used textbook definition frames AI as the design of rational agents — systems that act to achieve the best expected outcome given what they perceive and know. This framing is useful because it avoids arguing about whether a machine “really” thinks and focuses instead on measurable behaviour.
Artificial intelligence and robotics were among the research areas listed for CSE 2022 in December 2022. The concepts on this page are long-established foundations of the field. Where we mention developments after 2022, we describe them in general terms only; consult current peer-reviewed literature for the state of the art.
02Historical AI Research
- 1950Turing's “Computing Machinery and Intelligence”Proposed the imitation game as a way to discuss machine intelligence.
- 1956Dartmouth summer workshopThe term “artificial intelligence” became established as the name of a research field.
- 1970s–1980sExpert systemsRule-based systems encoded specialist knowledge; maintaining large rule bases proved difficult.
- 1990s–2000sProbabilistic and statistical methodsBayesian networks and statistical learning handled uncertainty more robustly.
- 2010sDeep learningLarge datasets, GPUs and improved training methods brought major gains in vision and speech.
- 2017 onwardTransformer architecturesAttention-based models became the basis of large language models used widely after 2022.
03Symbolic Artificial Intelligence
Symbolic AI represents knowledge as explicit symbols and rules, and reasons by manipulating them with logic. Its strengths are transparency and precision: every conclusion can be traced to the rules that produced it. Its weaknesses are brittleness and the “knowledge acquisition bottleneck” — the difficulty of writing down everything a system needs to know.
Symbolic methods remain important in theorem provers, planning systems, constraint solvers and formal verification, and research on combining them with learning (often called neuro-symbolic AI) is active.
04Search and Planning
Many AI problems can be framed as search through a space of states: find a sequence of actions that leads from an initial state to a goal. Uninformed strategies such as breadth-first search explore systematically; informed strategies use a heuristic estimate of distance to the goal.
The A* algorithm expands the node with the lowest value of
where g(n) is the cost so far and h(n) the heuristic estimate of the remaining cost. If h never overestimates the true cost (it is admissible), A* is guaranteed to find an optimal path. Planning extends search to actions with preconditions and effects, and game-playing algorithms such as minimax with alpha–beta pruning handle adversaries. See algorithms and data structures for the underlying graph search methods.
05Knowledge Representation
Knowledge representation studies how to encode facts about the world so that a system can reason with them. Approaches include propositional and first-order logic, semantic networks, frames, ontologies that define concepts and their relationships, and knowledge graphs that store entities and relations at large scale. Choosing a representation is a trade-off between expressiveness and the computational cost of reasoning.
06Intelligent Agents
Agents range from simple reflex agents, which map percepts directly to actions, to utility-based agents that weigh outcomes and learning agents that improve with experience. Environments differ in whether they are fully observable, deterministic, static or shared with other agents, and each property changes which methods are appropriate. Robotics is the study of physical agents, where perception is noisy and actions have real-world consequences.
07Machine Learning within AI
Machine learning is the branch of AI concerned with improving performance from data or experience rather than from hand-written rules. It has become the dominant approach in perception and language tasks, but it is a subset of AI, not a synonym for it. Learning paradigms, evaluation metrics and failure modes are covered in detail in machine learning applications.
08Natural Language Processing
Natural language processing (NLP) enables computers to analyse, understand and generate human language. Classic tasks include tokenization, part-of-speech tagging, parsing, named-entity recognition, machine translation, summarization and question answering.
The field moved from hand-written grammars, to statistical models trained on corpora, to neural models that learn vector representations of words and sentences. Language is ambiguous, context-dependent and culturally situated, so systems that appear fluent can still misunderstand meaning or produce confident but incorrect statements.
09Computer Vision
Computer vision extracts information from images and video: classifying what an image shows, detecting and locating objects, segmenting regions, estimating depth and tracking motion. Convolutional neural networks were central to progress in the 2010s because their structure suits spatially local patterns. Vision systems are sensitive to changes in lighting, viewpoint and context that humans handle easily, which is why testing under varied conditions is essential.
10Reasoning and Decision Making
Real decisions involve uncertainty. Probabilistic reasoning uses Bayes' theorem to update beliefs as evidence arrives; Bayesian networks represent dependencies among variables compactly. Decision theory combines probabilities with utilities to choose actions, and Markov decision processes model sequential decisions where each action affects future states — the formal basis of reinforcement learning.
11AI Evaluation
Evaluating AI is harder than it looks. Common practice compares systems on shared benchmark datasets, but a high benchmark score does not guarantee good behaviour in deployment. Sound evaluation asks:
- Is the test data genuinely separate from training data, with no leakage?
- Does the benchmark represent the conditions where the system will be used?
- Are results reported with variation across runs, not a single best score?
- How does the system fail, and how often, on difficult or unusual cases?
- Has it been compared fairly with simple baselines and with human performance where relevant?
12Responsible AI and Research Ethics
AI systems can affect people's access to jobs, credit, healthcare and information. Responsible AI research considers fairness across groups, transparency about how systems work, privacy of training data, safety and robustness, accountability for outcomes, and environmental cost. Practical tools include documenting datasets and models, auditing performance across demographic groups, involving affected communities, and obtaining ethical review for research with human participants.
Frameworks such as the NIST AI Risk Management Framework (2023) offer structured guidance for identifying and managing these risks.
13Applications and Limitations
| Application area | Typical use | Key limitation |
|---|---|---|
| Scientific discovery | Screening candidate molecules or materials | Predictions require experimental confirmation |
| Engineering | Fault detection, design optimization | Rare failures may be absent from training data |
| Language technology | Translation, summarization, search | Fluent output can be factually wrong |
| Robotics | Navigation, manipulation | Transfer from simulation to real world is difficult |
| Healthcare | Image analysis support | Performance can drop across hospitals and populations |
Further reading
- Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.
- Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460.
- Jurafsky, D., & Martin, J. H. (2009). Speech and Language Processing (2nd ed.). Pearson Prentice Hall.
- National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0) (NIST AI 100-1).