View some short excerpts from the film 2001: A Space Odyssey

• Scene 13: Crew members are interviewed on the news program "The World Tonight"
• Scene 15: Chess with HAL
• Scene 16: Sketches and Suspicions

HAL's capabilities

• Natural language production and understanding
• Common sense knowledge
• Planning
• Vision
• Emotion
• Game playing
• Learning

Present capabilities of Artificial Intelligence (AI)

• A variety of systems exhibit impressive task-specific competencies
  • NetTalk learned to pronounce English text
  • NavLab learned to drive a car in the city and on the highway
  • Deep Blue beat the human World Champion in chess
• But successes have been in narrow domains
• Current systems lack the integration, generality and adaptiveness seen in HAL

Forecast looks bleak for achieving HAL's level of intelligence

• The most interesting applications for AI (exploring other planets, cleaning up chemical spills, etc.) involve much more complex domains
• Broadening the capabilities of an AI system will place an increasing burden on the human designers
• This does not bode well for achieving goals of integration, adaptiveness and generality

A possible solution

• Allow a process akin to natural selection to evolve AI systems
• It has been predicted by a number of prominent AI researchers that the evolutionary approach will soon supersede design by hand
• Offers a means for overcoming the bottleneck of the human designer

Common components of simulated evolution

• Operate on a population of individuals whose genes are typically represented as a string of real-valued parameters
• Evaluate each individual according to a fitness function
• Use a selection mechanism to simulate survival of the fittest
• Use nature-inspired operators to create children from parents, such as mutation and crossover
• Use a replacement policy to keep population size fixed
• The goal is to find a population that optimizes the fitness function

Algorithm for one method: Evolutionary Strategies

1. parents = 5, children = 50, time = 0, limit = 500
2. randomly initialize parameters of first generation parents
3. repeat
(a) create 10 children from each parent through mutation
(b) determine the fitness of each child
(c) replace parents with 5 fittest children
(d) time = time + 1
4. until best individual is good enough or time > limit

Human designer's role in evolutionary approach

• Choosing the appropriate parameters to be evolved
• Choosing a fitness function
• Determining the evolutionary framework

Once this evolutionary approach is set in motion, the rest of the design process is automated

What should be evolved to create robust AI systems?

• Complete programs in a high-level language such as C or LISP
• Parameter settings for a hand-designed system
• A set of if-then rules
• Artificial neural networks

My research focuses on evolving artificial neural networks to control robots

Advantages

• Provide real-time, noise-tolerant performance
• Operate at a low level so are less affected by designer bias
• Offer more potential to discover new modes of behavior
• May lead to a deeper understanding of how perception and action are linked

Disadvantages

• Can be slower to develop good results
• Can be difficult to achieve high levels of complexity

Characteristics of biological neural networks

• Human brains contain approximately 10^10 neurons
• Neurons are densely interconnected with up to 10^5 connections per neuron
• Connections can be excitatory or inhibitory
• When neurons are excited their firing rate increases
• Learning involves modifying synapses
• Elimination and addition of connections can occur

Characteristics of artificial neural networks

• Artificial networks typically contain less than 10^4 neurons
• Artificial neurons have up to 10^3 connections
• Connections are represented as real-valued weights which can be positive or negative
• Firing rate is represented as a real-valued activation level
• Learning involves modification of the weights

A neural network for robot control

Contains 9 neurons and 18 weights

Khepera robot

Task to be learned: Navigate through a labyrinth

Possible fitness functions

• Explicit: Maintain equal distance between walls, avoid collisions, and maximize speed
• Intermediate: Maximize speed, straightness, and distance from walls
• Implicit: Maximize ground covered

Results

• Intermediate level fitness functions are the most successful
• Allow both the environment and the fitness function to shape behavior

Conclusions

• The evolutionary approach is a promising alternative for developing artificial intelligence
• Computer speed and memory capacity continue to increase exponentially
• Ray Kurzweil predicts that a $1000 personal computer will match the computing speed and capacity of the human brain by the year 2020
• Hans Moravec predicts that a humanoid robot will be viable by by the year 2030