Neurale netværk og spillet fire på stribe: Neural Networks and The Game Four In a Row

Daniel Borup, Jakob R Hestbæk, Jesper S Larsen, Jon Nielsen, Jonas Rømer & Arne Schroppe

Studenteropgave: Semesterprojekt

Abstrakt

Abstrakt (Dansk) Projektets formål er at undersøge kunstige neurale netværk (neurale netværk) og regelbaserede systemer (ekspertsystemer) som modeller for kunstig intelligens, og nå frem til en forståelse for potentiale, styrker og svagheder ved disse to modeller. Med udgangspunkt i eksisterende teori på området, implementerer vi modellerne og eksperimenterer med dem. Eksperimenterne udføres med spillet fire på stribe. Vi fandt ud af at et regelbaseret system med få regler kan vinde fire på stribe mod menneskelige spillere i en del af spillene. En ulempe ved ekspertsystemet er at man kan lære at gennemskue dets regler. Det viste sig at være svært at træne det neurale netværk op til et niveau hvor det kan slå menneskelige spillere, men efter træning med serier af træk fra et ekspertsystem som vandt over et andet, nåede det neurale netværk en succesrate på 70 % mod ekspert systemet. Det kunne indikere at det neurale netværk har lært en strategi som virker mod ekspertsystemet. Samlet set kan vi konkludere at et neuralt netværk formentlig vil kunne trænes op til vinde spillet fire på stribe med høj succesrate mod menneskelige spillere, men det kræver yderligere arbejde med topologien af det neurale netværk, samt bedre træningsdata. Abstract (English) This project takes a closer look at artificial neural networks (neural networks) and rule based systems (expert systems) as models for artificial intelligence, with the purpose of reaching an understanding of the potential strengths and weaknesses of the two systems. With basis in existing theories, we will implement and experiment with these models. We will use the game of 4-in-a-row as backbone in our experiments. We would learn that an expert system with just a few rules, can often win in 4-in-a-row over a human player. One disadvantage of expert systems is that it is easily beaten, if one is able to learn the rules on which it is based, and able to exploit potential weaknesses or flaws. It should however prove difficult to train the neural network to a level where it would be able to beat an average human player. But after letting the neural network learn by watching two expert systems playing each other, the neural network reach a success rate of 70 % against the expert system. This could indicate that the neural network has learned a strategy that works against the expert system. Looking at the big picture, we can conclude that it would be possible to teach a neural network to win over a human player in 4-in-a-row with a high success rate. But it would require a lot more work with the topology of neural networks, as well as better training.

UddannelserBasis - Naturvidenskabelig Bacheloruddannelse, (Bachelor uddannelse) Basis
SprogDansk
Udgivelsesdato1 jun. 2005
VejledereMads Rosendahl

Emneord

  • neurale netværk neural networks neural net