Introduction to Artificial Intelligence Flashcards
Thinking humanly
Cognitive science approach to make humans think humanly
Thinking rationally
Rule-based thinking using algorithmic thinking
Acting humanly
Passing the Turing test
Definition of intelligence
- Carry out abstract thinking
- Think rationally
- Play chess
- Score high at an IQ test
- Ability to learn from past experiences.
Reductionism
- Break down the problem into manageable pieces
- Study each piece separately
- Put the pieces together again to understand the whole
- TOP-DOWN perspective
Holism
Argues that reductionism isn’t possible (i.e. to break down the problem) because we lose important pieces not present in any one part. These are called emergent properties.
Studies simple but complete systems. BOTTOM-DOWN perspective.
Analytical approach
- Reductionistic approach
- Primary method for empirical analysis
- Hypothesis
- Test
- Analyze results
Synthetic approach
Understand by building a model of the problem or structure. AI and cognitive science is mostly synthetic. Synthetic and analytical approaches are complementary, not contradictory.
Microworlds
A subset of the world for which it is easy to create a domain ontology
Agents
Percieve things and act. Humans use eyes, ears, touch and more to perceive it’s environment and then act on it.
Percept sequence
Everything an agent has ever perceived.
Percept
What an agent perceives atm
Rational agent
An agent that at all times do the actions with highest expected outcome-value. I.e. maximize it’s performance measure, so it depends on what you define as measure. Also, given the evidence provided by the percept sequence and whatever built-in knowledge the agent has.
The rationality of an agent depends on:
- Possible actions
- Performance measure
- Percept sequence
- Built-in knowledge
What is the Task environment?
PEAS (Performance, Environment, Actuators, Sensors)
Types of Task environments
- Partly vs. fully observable
- Static vs. dynamic
- Discrete vs. continuous
- Episodic vs. sequential
- Deterministic vs. stochastic
- Single vs. multi agent
- Known vs. unknown
Differences between fully and partly observable environment
Fully observable: The sensors detect all aspects relevant to the choice of action. Depends on the sensors, environment and performance measurement.
Partly observable: Most real-life tasks
Unobservable: No sensors
Differences between static and dynamic environment
Dynamic: Can change as the agent makes a choice. Ex: cooking agent can burn the food while it cooks.
Semi-dynamic: The environment doesn’t change but the score does.
Static: No change in environment no matter the time it takes to complete the task. Ex: Sudoku-solver.
Discrete vs continuous environment
Discrete: Can be divided into a finite number of states.
Continuous: Can not. There are several reasons for this:
1. The environment can change continuously (weather)
2. The agent can take input continuously (infra-red sensor), and or
3. The agent acts in a continuous matter (driving)
Note: One can often model discrete states as continuous and vice versa.
Episodic vs sequential environment
Episodic: Atomic happenings (all are individual with no parallell connections)
Sequential: One decision can/will affect later incidents and decisions. The agent can still ignore previous actions, though (reactive)
Deterministic vs Stochastic environment
Deterministic: Where the next state can be completely determined by the previous one.
Stochastic: Where it doesn’t only matter what previous states one has taken. That is, it’s not random, it still builds on from previous actions.
Single vs multiagent environments
Single: One agent acts alone.
Multi: More agents. That is, if there’s another element with a performance measure, which depends on the actions of the other agent. Then, we have more than one agent.
Different types of multiagent environments
Cooperative: The actions of one agent increases the betterment for all agents (fleet of vacuum robots)
Competitive: All agents cannot be bettered simultaneously.
Known vs unknown environment
All probabilities of the outcomes of actions the agent makes are known in a known environment.
All agent types
- Simple vs model-based
- Reactive
- Goal based
- Utility based
- Learning agents
Simple agent and Model-based agent
Doesn’t store an internal representation of the world. React exclusively to the percepts. Good for fully observable scenarios.
Model-based keep a representation of the world, and uses it to determine better actions together with percept-
Reactive agents
Acts fully on the current precept it receives. It ignores the history. Uses if-then rules.
Goal-based agents
They take actions based on completing some goal.
Algorithm:
Sense –> Create model of the environment –> look for a sequence of actions that will complete the goal –> Execute the first action.
If it operates in a dynamic environment it needs to ask “what if I do this, then what?”. That is, it models both the environment and the goal.
Utility-based agents
Chooses actions based on a utility function, which weights more than one factor. For example, an autonomous car we want to weight in safety, speed, gas consumption etc.
Different representations of the world
- Atomic (Nodes in a graph)
- Factored (variables and combinations of them)
- Structured (states and relationships between them)
Difference between traditional and declarative problemsolving
Traditional: User –> program (how?) –> Computing environment
Declarative: User –> knowledge base (what?) –> Solver (how!) –> Computing environment
Workflow for traditional programming
Problem –> (Programming) –> Program –> (Solve) –> Output –> (Interpreting) –> Solution
Workflow for declarative programming
Problem –> (Modeling) –> Representation –> (Solving) –> Output –> (Interpreting) –> Solution
What is declarative programming used for? Traditional?
It’s used to support in decision making.
Logic
Formal language for interpreting happenings so conclusions can be drawn.
What are the two components of logic?
Syntax: Defines the sentences
Semantics: Defines the meaning of the sentences
Monotonic logic vs non-monotonic
Monotonic: The truth of axioms doesn’t cahnega s new information is added
Non-monotonic: It might change (‘birds typically flies and Birdy is a bird, Birdy presumable flies)
Satisfiable formula
IFF it’s True for at least one interpretation of the expression
Valid formula
IFF True for all interpretations
Unsatisfiable formula
Not True for any interpretation
Model
every interpretation that satisfies a formula is a model of that formula
Truth table
Look up online and learn to create….
Reasoning
Inference - the process of answering a question
Entailment
One thing follows from another.
Formal properties of reasoning
- Soundness - the initial reason for counting the logic as desirable
- Completeness - all true states are provable
Completeness is important in business to manage “business rules”
Describe weak AI and the differences to strong AI
Focused on small tasks and tries to get good at only that, strong AI is like the human mind, but better, and can do anything a human mind can and more
Describe Artificial general intelligence
Equally smart as the human mind