Artificial intelligence - Goals Subfields Core Concepts and Approaches

Goals and Subfields of Artificial Intelligence Introduction Artificial intelligence encompasses a broad range of techniques and approaches aimed at creating machines capable of performing tasks that typically require human intelligence. These goals and subfields can be understood as two complementary perspectives: the specific research goals that drive AI development, and the distinct technical subfields that pursue those goals using different methodologies. Understanding both perspectives is essential for grasping how AI systems are built and evaluated. Traditional Research Goals AI research is organized around seven major goals, each representing a fundamental capability that intelligent systems should possess. Reasoning enables machines to draw logical conclusions from given information. A reasoning system takes facts as input and applies logical rules to derive new conclusions—much like how humans work through a mathematical proof or a legal argument. Knowledge representation addresses how to store facts, concepts, and relationships in a machine-readable form. This is more challenging than it might appear: encoding "dogs are animals" or "water boils at 100°C at sea level" requires formal structures that a computer can process and reason with effectively. Planning involves generating sequences of actions that, when executed, will achieve specified goals. For example, a robot that needs to move an object from point A to point B must first plan a sequence of movements—pick up object, move forward, place object down. Learning focuses on improving a system's performance on tasks through experience or data. Rather than programming every behavior explicitly, learning systems acquire knowledge by processing examples. Natural language processing (NLP) enables computers to read, write, and communicate in human languages. This includes understanding the meaning of text and generating coherent responses, which is surprisingly complex given the ambiguity inherent in human language. Perception gives AI systems the ability to interpret sensory inputs such as images from cameras or audio from microphones. Perception bridges the gap between raw sensor data and meaningful information about the world. Robotics involves applying AI techniques to control physical agents that can move and manipulate objects in the real world. Robotics combines perception, planning, and control to enable autonomous behavior. Long-Term Aspirations: General Intelligence Beyond these specific research goals lies a more ambitious aspiration: general intelligence. A system with general intelligence could perform any intellectual task that a human can perform, at equal or greater proficiency. This contrasts sharply with the specialized systems we have today, which excel at narrow, well-defined problems. General intelligence remains largely theoretical—we do not yet know how to achieve it, and some researchers debate whether it's even possible. Subfields of Artificial Intelligence Reasoning and Problem-Solving Early AI research in the 1950s and 1960s attempted to solve reasoning problems by simulating step-by-step logical deduction. These algorithms would explore possible inferences one at a time, mimicking how a human mathematician might work through a proof. A major limitation of this approach emerged quickly: many real-world reasoning problems suffer from combinatorial explosion. As the problem size grows, the number of possible logical steps and conclusions grows exponentially, making computation infeasible. For instance, exploring every possible move in chess becomes impossible beyond a few moves ahead. By the late 1980s, researchers incorporated probability and economics into reasoning systems. Instead of treating knowledge as absolute truth, these systems assigned probabilities to uncertain facts and used expected value calculations to make decisions. This probabilistic approach proved more practical for handling real-world uncertainty. Knowledge Representation and Engineering At the foundation of any intelligent system lies a knowledge base: a collection of information stored in machine-readable form. A knowledge base might contain facts like "Paris is the capital of France" or rules like "if something is a dog, then it is a mammal." An ontology provides the conceptual structure for a domain. It defines the objects, relationships, concepts, and properties that matter in a particular field. For instance, a medical ontology would define diseases, treatments, symptoms, and the relationships between them. A key challenge in knowledge representation is capturing commonsense knowledge—the vast body of everyday facts that humans take for granted. Humans know that wet floors are slippery, that people sleep at night, or that if you drop an object it falls. Encoding these millions of commonsense facts has proven surprisingly difficult. Another major hurdle is knowledge acquisition: actually obtaining and organizing the knowledge needed for a system. This often requires significant manual effort from domain experts, making it time-consuming and expensive. Planning and Decision-Making An autonomous agent is a system that perceives its environment through sensors and takes actions through effectors to achieve goals or satisfy preferences. A self-driving car is an agent—it perceives the road through cameras and lidar, and acts through acceleration, braking, and steering. Rational agents assign numeric values called utilities to different outcomes, reflecting how desirable each outcome is. An agent prefers outcomes with higher utility. The expected utility of an action is calculated as: $$\text{Expected Utility} = \sum{\text{outcomes}} P(\text{outcome} \mid \text{action}) \times U(\text{outcome})$$ This formula says that the expected utility of an action equals the sum of the utilities of possible outcomes, weighted by the probability that each outcome will occur given that action. Classical planning assumes that actions have deterministic effects—when you execute an action, you know exactly what will happen. However, real-world planning rarely works this way. Uncertainty is everywhere: a robot might intend to grasp an object and miss, or a weather prediction might be slightly inaccurate. Real planning systems must account for partial observability (not knowing the current state perfectly) and stochasticity (randomness in action effects). Markov decision processes (MDPs) provide a formal model for handling these uncertainties. An MDP defines a set of states, a set of possible actions, a probability distribution over state transitions (what state you'll be in after each action), rewards for reaching certain states, and a policy (a strategy that prescribes which action to take in each state). MDPs are central to decision-making under uncertainty. Game theory extends decision-making to situations with multiple interacting agents, each with their own goals. Game theory helps us understand rational behavior when the outcome depends on what other agents do—like in chess, poker, or economic negotiations. Machine Learning Machine learning studies programs that automatically improve their performance on tasks through experience rather than explicit programming. Instead of a programmer writing rules, a machine learning system learns those rules from data. There are three primary learning paradigms, distinguished by the type of feedback available: Supervised learning uses labeled training data—examples where the correct answer is known. A supervised learning system learns to predict outcomes from inputs. This includes: Classification: predicting which category an input belongs to (e.g., "Is this email spam or not spam?") Regression: predicting a numeric value (e.g., "What will the house price be?") Unsupervised learning discovers patterns in data without labeled examples. When you have a large collection of unlabeled data, unsupervised learning algorithms might discover that the data naturally clusters into distinct groups or that certain features tend to vary together. This is useful when you don't yet know what patterns to look for. Reinforcement learning is fundamentally different: an agent learns by interacting with an environment, receiving rewards for good actions and penalties for poor ones. Over time, the agent learns which actions lead to high total reward. This mimics how animals learn through trial and error. Transfer learning applies knowledge gained from solving one problem to a different problem. If you've trained a system to recognize dogs in images, you might transfer that knowledge to recognize cats by reusing the learned features and adjusting only the final classification layer. Deep learning has become the dominant approach in modern AI. Deep learning systems process inputs through artificial neural networks with many layers. Each layer learns to detect increasingly complex patterns. Deep learning has driven most recent breakthroughs in vision, language, and game-playing. Natural Language Processing Natural language processing tackles the challenge of making computers understand and produce human language. Key tasks include: Speech recognition: converting audio into text Speech synthesis: converting text into natural-sounding audio Machine translation: translating between languages Information extraction: identifying structured facts from unstructured text Information retrieval: finding relevant documents in response to queries Question answering: providing direct answers to natural language questions Modern NLP relies on several powerful techniques. Word embeddings represent words as vectors in a high-dimensional space, where semantically similar words are positioned near each other. Transformer architectures use an attention mechanism that allows the model to focus on the most relevant parts of the input. Large pre-trained language models—systems trained on vast amounts of text—can be fine-tuned for specific tasks, making them remarkably versatile. Machine Perception Perception uses sensor data from cameras, microphones, lidar (laser scanning), radar, and other sensors to infer properties of the world. Perception is the bridge between raw sensory information and high-level understanding. Computer vision analyzes visual input for tasks including: Image classification: assigning a label to an image ("This is a dog") Object detection: locating and identifying multiple objects in an image Facial recognition: identifying who is in an image Tracking: following a moving object across a video Speech recognition converts audio signals into text, enabling voice interfaces and transcription systems. Evaluating Approaches to AI Symbolic Artificial Intelligence The earliest AI research took a symbolic approach, assuming that intelligence could be achieved by manipulating symbols according to explicit rules. Symbolic systems attempted to simulate high-level reasoning tasks like algebra, legal reasoning, and mathematics. This approach was motivated by the physical symbol systems hypothesis, proposed by computer scientists Allen Newell and Herbert Simon. They claimed that a physical system—like a computer—that can manipulate symbols has the necessary and sufficient means for general intelligent action. In other words, if you can encode knowledge as symbols and write rules for manipulating those symbols, you can achieve intelligence. However, symbolic AI encountered a fundamental problem: it excelled at formal, logical tasks but struggled with abilities that humans find effortless. Learning from experience, recognizing objects in images, and applying common sense all proved far harder to encode as symbolic rules than anticipated. This mismatch between the difficulty of "intelligent" reasoning tasks and the ease of basic perception and instinct is known as Moravec's paradox. <extrainfo> Moravec's Paradox and Human Expertise Moravec's paradox observes a striking counterintuition: high-level "intelligent" tasks like playing chess or solving mathematical proofs are relatively easy for AI, while low-level "instinctive" tasks like recognizing faces or walking on uneven ground are extraordinarily difficult. Philosopher Hubert Dreyfus provided an explanation for this paradox. He argued that human expertise does not rely primarily on explicit symbolic manipulation—that is, consciously following rules. Instead, expertise relies on unconscious instinct and a "feel" for the situation developed through experience. A chess master, for instance, does not consciously calculate the value of every possible move; experienced players recognize patterns intuitively. A mother does not calculate the physics of balance to walk upright; she does it naturally. These embodied, instinctive skills turned out to be harder to automate than abstract reasoning. </extrainfo> Methodological Philosophies in AI Beyond symbolic versus non-symbolic approaches, AI researchers hold different philosophical views about how intelligence should be pursued. <extrainfo> Neat versus Scruffy Approaches The "neat" approach seeks to describe intelligent behavior using simple, elegant, unified principles—such as logic, optimization theory, or artificial neural networks. Neats believe that intelligence has an underlying elegant structure that can be discovered and formalized. The "scruffy" approach, by contrast, believes that intelligence is too complex to reduce to simple principles. Instead, intelligence requires solving a large variety of unrelated problems using different techniques. Scruffies trust incremental testing and practical experimentation over theoretical rigor. They are willing to apply different solutions to different problems rather than seeking one unified approach. Soft versus Hard Computing Hard computing seeks provably correct or optimal solutions using classical mathematics and logic. Hard computing provides guarantees: if you solve a problem with hard computing methods, you know the answer is correct. However, hard computing is often intractable for important real-world problems. Finding the optimal solution might require checking an infeasible number of possibilities. Soft computing includes techniques like genetic algorithms, fuzzy logic, and neural networks that tolerate imprecision, uncertainty, partial truth, and approximation. Soft computing is practical for complex problems where exact solutions are impossible to compute. The tradeoff is that you get a good solution, but not necessarily the best one. </extrainfo> Narrow versus General Artificial Intelligence A crucial distinction exists between two levels of AI capability: Narrow artificial intelligence (also called weak AI) focuses on solving specific, well-defined problems with tailored solutions. All current AI systems are narrow. A language model excels at predicting text but cannot navigate a city. A chess engine plays chess superbly but cannot diagnose diseases. Modern AI is built by assembling many narrow specialists. General artificial intelligence (also called strong AI) seeks to create systems that can perform any intellectual task that a human can perform. A generally intelligent system would be flexible, capable of learning new domains, and able to transfer knowledge across disparate problems. Despite decades of research, general intelligence remains elusive. The definition itself is slippery—what exactly should we measure to know if a system is generally intelligent?—and whether it is even possible remains debated among researchers. Core Concepts for Review To solidify your understanding, here is a summary of essential terminology: Agents and goals: An agent perceives its environment through sensors and acts upon it through effectors to achieve defined goals. Learning paradigms: Supervised learning trains on labeled examples; unsupervised learning discovers patterns without labels; reinforcement learning learns through trial-and-error interaction with an environment. Knowledge representations: Symbolic representations (semantic networks, frames, logic statements) encode explicit knowledge. Sub-symbolic representations (neural network weights) capture implicit knowledge learned from data. Evaluation metrics: In task-centric definitions of AI intelligence, the difficulty of a problem and the system's performance on that problem serve as direct measures of intelligence. Ethical and societal considerations: Bias in algorithmic decision-making can create unfair outcomes for different demographic groups. Transparency and explainability—understanding why a system made a particular decision—are essential for responsible AI deployment.