COGSCI1: Lecture 12: Architectures, Modules, and Dynamical Systems Theory

How can we build an intelligent agent?/Why did we build them initially?

we made them initially in order to create a model that can test hypotheses & reality

What are artificial agents?/What features do they have?

• autonomous & don’t need to be guided at any point

• need to be able to sense the world in order to interact with it

• need to be able to effect the world

Problem #1 (easy): how to act immediately in predictable environments

• most basic thing you can do to be autonomous to some degree

• this architecture is known as the simple reflex agent

Simple Reflex Agent

• info comes in through the sensors and goes out through the actuators

• inside the model, it is evaluating info

Simple Reflex Agent: sensors

sensors tell us what the world is like now (are we on the way to what we should be doing or not)

• using this you can choose what action should be done, and that sends information to the actuators

Simple Reflex Agent: actions

• don’t have many choices w/ the actions:

  • there are production statements (if-then rules)

  • not great w/ unpredictable/new environments

    • it is a very basic design

Problem #2 (harder): how to act flexibly when multiple actions exist

• need to select out of multiple options

  • still has the same architecture as simple reflex in sense of where are we now vs. where do we want to be

• now also an ability to simulate what happens if you choose action A vs. action b (needs to perform some kind of estimation)

  • to simulate you need a model

• this architecture is the goal-based agent

Goal-Based Agent: choosing

• knows the state of the world right now, and has a model of how the world evolves & what their actions do

  • use all of this to figure out what happens if they choose action A

• to choose what action they should do now, they also need to incorporate the goals

• is fine as long as you don’t encounter things outside of the range/environment, but still flexible enough to deal with some different situations differently - you have some context

Problem #3 (even harder): how to adapt and improve without being told what to expect

• need an agent that is capable of new things

• this architecture is the learning agent

Learning Agent: