Presentation is loading. Please wait.

Presentation is loading. Please wait.

1 Chapter 13 Artificial Life: Learning through Emergent Behavior.

Similar presentations


Presentation on theme: "1 Chapter 13 Artificial Life: Learning through Emergent Behavior."— Presentation transcript:

1 1 Chapter 13 Artificial Life: Learning through Emergent Behavior

2 2 Chapter 13 Contents (1) l What is life? l Emergent Behavior l Finite State Automata l Conway’s Life l One-Dimensional Cellular Automata l Self-Reproducing Systems

3 3 Chapter 13 Contents (2) l Evolution l Evolution Strategies l Genetic Programming l Evolutionary Programming l Classifier Systems l Artificial Immune Systems

4 4 What is life? l What are the defining features of life? nself-reproduction nability to evolve by Darwinian natural selection nresponse to stimuli nability to die ngrowth or expansion l Not all living things obey these rules, and some things that are not alive do. l Defining life is very difficult!

5 5 Emergent Behavior l The idea that complex behavior emerges from simple rules. l Is seen in systems such as CYC, but is particularly prevalent in systems based on evolutionary methods, such as genetic algorithms. l Example: nBoids – simulations of birds given very simple rules about how to fly. nAutomatically flew in such a way as to avoid large obstacles, without being taught explicitly how to do so.

6 6 Finite State Automata l FSA: a machine with a finite number of states. l The FSA is given inputs, which result in transitions between states. l Some states are accepting, meaning the FSA is saying “Yes”. l Other states are rejecting. l In this example there are two possible input characters – a and b, and two states, 1 and 2. l It will finish in state 1 (the accepting state) if the input has an even number of a’s.

7 7 Conway’s Life (1) l A two dimensional cellular automaton. l A two dimensional grid of cells, each of which can be alive or dead. l A set of rules determines how the cells will change from one generation to the next: 1.A dead cell will come to life if it has three living neighbors. 2.A living cell with two or three living neighbors, will stay alive. 3.A living cell with fewer than two living neighbors will die of loneliness. 4.A living cell with more than three living neighbors will die of overcrowding.

8 8 Conway’s Life (2) l Surprisingly complex behavior can sometimes emerge from these simple rules. l This diagram shows a successive sequence of generations of Conway’s Life. l This pattern is known as a glider. l There is also a pattern known as a glider gun which constantly fires out gliders.

9 9 One-Dimensional Cellular Automata (1) l A single line of cells. Each cell thus has two immediate neighbors. l It is usual to have rules that take into account the cells on either side of the immediate neighbors as well. l Usually, the cell itself is also taken into account, meaning that each cell’s future is determined by 5 cells. l 1-D Cellular Automata often use totalistic rules, meaning that the number of living cells out of the 5 is all that determines the cell’s state in the next generation.

10 10 One-Dimensional Cellular Automata (2) l Hence, there are 32 possible rule sets. One such set of rules might be: 1 2345 0 0111 l This says a cell can only survive if it has two, three or four neighbors. l This rule can be seen applied in five generations in the following diagram:

11 11 Self-Reproducing Systems l Von Neumann proposed a self-reproducing system based on cellular automata. l Langton invented loops: nEach loop consists of 94 cells. nContains all the information that is needed to reproduce itself. l Ultimately we may have robots that can obtain the raw materials necessary to produce new versions of themselves. l This would be useful for exploring other planets.

12 12 Evolution l The changes in cellular automata involve single-step selection, while evolutionary systems involve cumulative selection (Dawkins, 1991). l Survival of the fittest means that creatures that are fit are more likely to reproduce than those that are less fit. l This idea is modeled exactly in systems such as genetic algorithms.

13 13 Evolution Strategies l Similar to hill-climbing. l A set of numeric parameters is varied from generation to generation by making normally distributed changes to the values. l If the offspring is a better solution than the parent, then the process repeats from the offspring. l Otherwise, the offspring is rejected, and a new attempt is made. l This is asexual reproduction – a single parent produces a single offspring.

14 14 Genetic Programming l A method used to evolve LISP S-expressions. l The S-expressions are represented as trees. l A random set of expressions is generated, and the “fittest” individuals reproduce to produce the next generation. l Mutation and crossover are used (see chapter 14). l Diagram shows an example of a tree representation of an S-expression.

15 15 Evolutionary Programming l Evolves finite state automata to solve the problem of identifying the next item in a sequence: a 1, a 2, a 3, a 4, a 5, …, a n. l A new generation of FSAs is made by applying a set of mutation operators: l 1) Changing an output symbol l 2) Changing a state transition l 3) Adding a state l 4) Deleting a state l 5) Changing the initial state l The success of an FSA is determined by seeing how well it predicts the existing sequence.

16 16 Classifier Systems (1) l An evolutionary expert system. l Has the following components: nDetectors which receive inputs from the environment. nEffectors which send outputs and carry out actions. nA rule system, which consists of a population of classifiers. Each rule has a measure of fitness. nDetectors to determine how well the system is performing. nA bucket brigade algorithm for assigning credit and blame to classifiers. nA procedure for reproducing classifiers by application of a set of genetic operators. nA set of message lists – for input, output and internal messages.

17 17 Classifier Systems (2) l The system uses classifiers to determine what outputs to produce, or what actions to carry out depending on the inputs from the environment. l A classifier has the following form: (c 1, c 2, c 3, c 4, c 5 ) -> M, f l c 1 … c 5 are the input variables; M is the output or action and f is the fitness of the classifier. l An example classifier might be: (4, 2, *, 1, *) -> A2, 9.1 l * represents any input. l This rule states that an input that has 4, 2 and 1 in the first second and fourth positions is classified as A2, and that the classifier has a fitness value of 9.1.

18 18 Classifier Systems (3) l When a system has classified an input, a new generation of classifiers is produced by allowing the classifiers that provided the best classifications to reproduce. l A bucket-brigade algorithm is used to assign credit (or blame) to the classifiers. l Classifier systems can be used to solve a number of problems, including playing games and enabling a virtual robot to explore a virtual terrain.

19 19 Artificial Immune Systems l Systems modeled on the immune systems in humans and other biological creatures. l Used in anti-virus systems, for example. l Also applied in computer security, for solving combinatorial problems, and for machine learning problems.


Download ppt "1 Chapter 13 Artificial Life: Learning through Emergent Behavior."

Similar presentations


Ads by Google