Presentation on theme: "Computational Intelligence: Some Insights from Board Games Fernand Gobet Brunel University."— Presentation transcript:
Computational Intelligence: Some Insights from Board Games Fernand Gobet Brunel University
Board Games Seems a good choice Both humans and machines are good at them But with different approaches Has attracted the interest of some of the giants of computer science: Alan Turing Claude Shannon Herbert Simon John Holland
Brute Force May Lead to Intelligence, after All Humans, including world champions, search highly selectively Kasparov probably searches at most 10 positions in a minute In 1997, Deep Blue beat Kasparov using massive search (200 mio positions per sec) But similar approaches have failed in Go
Is Parsimony Critical? Research into chess suggests that grandmasters have hundreds of thousands of perceptual patterns and rules With algorithms such as C4.5, decision trees with more than dozens of nodes are seen as too large
How Much Knowledge? Human masters typically use a lot of knowledge Acquired by instruction, study, and experience Many computational systems use little domain-specific knowledge Bottom up induction of knowledge (rules, clusters, etc.)
Problem-Based Learning Humans often learn while solving problems -- e.g., playing games Many computational techniques passively learn by accessing information in databases Does this affect the quality of what is being learnt? Later performance?
Bounded Rationality and Optimality Humans typically learn and behave in a sub-optimal way--but in real-time Their knowledge is not structured optimally (e.g., presence of contradictions) Similar flavour with biological evolution and Microsoft software Many CI techniques aim at optimal learning Is this realistic?