I-700-ABMS AGENT-BASED MODELING & SIMULATION

Einingalíkön og hermun

 

4. Introduction to Models of Individual Minds: Knowledge & Skill

 

 

 

 

 

4.1
Knowledge Acquisition = Learning
  Definition of learning (A.I.) A system is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measure by P, improves with experience E.
  The two main types of learning Knowing WHAT and knowing HOW
Also called "knowledge" and "skill", respectively and "declarative knowledge" and "procedural knowledge" ("know-how")
  Humans are typically seen as goal-directed agents who actively seek information e.g. Piaget, 1978; Vygotsky, 1978
  Knowledge must be constructed from existing knowledge You always understand new things in context to what you knew before - there is no such thing as learning from scratch
     

 

 

 

 

 

 

 

 

 

 

 

 

4.2
Artificial Intelligence
  Primary motivation To get a machine to learn
  Different goal

This focus is different from what we will be talking about, because individuals in our system do not have to learn to discriminate among perceptual stimuli or learn to do their job, they only have to simulate the acquisition of knowledge.

We will take ideas from AI; however, ABMS does not share the main goal of AI.

  Main types of machine learning
  • Concept learning. Learning which instances belong to which class. Discovering class (generalization) based on instances.
  • Decision tree learning. One of the most widely used methods for inductive inference, the process of reaching a general conclusion from specific examples. Good tutorial.
  • Artificial Neural Networks (ANNs). Attemtp to simulate real neurons.
  • Genetic algorithms. Inspired by how genes work.
  • Bayesian (statistical) learning / Belief Networks. Probabilistic approach to inference. Evidence / observations used to infer the probability that a statement may be true.
  • Reinforcement learning. Only learning mechanism that is "on-the-job" and in this way starts to approach human-like learning. However, current implementations are far from the power of human and animal learning.
  Useful methods In general the work in A.I. is composed of various techniques. The ones most relevant to us are Bayesian and Decision trees.
     

 

 

 

 

 

 

 

 

 

 

 

 

 

4.3
Psychological Models & Studies
  Humans: Many types of learning
  • Classical conditioning (like Pavlov's dogs)
  • Operant conditioning (like Skinner's rats)
  • Imitation learning (like apes in experiments)
  • Language learning (like no other animal)
  • Transfer of learning
  There is no universal function that describes skill acquisition First rule: No single learning curve fits all - depends on the topic.
Most common though: logarithmic (i.e. improvement decreases logarithmically with practice)
  Main characteristic of Logarithmic curves Each unit of practice produces smaller and smaller improvements
  Easy tasks lead to ceiling effects i.e. an inability to improve further after a certain amount of training
  Highly difficult tasks can be improved indefinitely through training There seems to be no theoretical limit to human memory or learning capabilities.
  Transfer of training When you learn something the knowledge is typically more general than just applying to the context you learned it for. When you do a related task some of the knowledge will transfer to the new task.
  Negative transfer of training If the new task is very similar but requires reversing some of the routines you learned, we call it negative transfer of training - you may in fact be slower at learning something that is very similar to something you already know than something very different.
  Rummelhart and Norman (1978): 3 stages of learning
  1. acquisition of facts in declarative memory
  2. initial acquisition of procedures in procedural memory
  3. tuning: modification of procedures to enhance reliability and efficiency