This document summarizes key developments in artificial intelligence between 1990-2010, including:
- Hybrid agent architectures that combine reflex components with knowledge-based reasoning. Examples include Soar and Theo.
- Real-time AI techniques like anytime algorithms and bounded rationality that allow agents to operate under resource constraints.
- Multi-agent systems research on cooperation, competition, emergence and rationality in decentralized groups.
- Modern successes like autonomous vehicles, planning systems, language processing, game playing, classification/prediction, logistics and robotics. Neurological simulation through projects like Blue Brain.
3. Hybrid Architectures
0 Hybrid Architectures: combining the various agent
architectures.
0 E.g. combining reflex components with knowledge‐
based components.
0 Compilation: convert declarative information at the
deliberate level into more efficient representations,
eventually reaching the reflex level.
0 Soar (1987), Theo (1990).
0 Solutions to new problems are derived by deliberation,
but are compiled and stored at the reflex level for
solving future occurrences of similar problems.
6. Multiagent Systems
0 Cooperation
0 Competition
0 Swarm intelligence: performance measure applied
to collective behavior.
0 Decentralized representation
0 Emergent behavior
0 Weak emergence: the qualities of the system are
reducible to the system's constituent parts.
0 Strong emergence: e.g. qualia.
0 The concepts of utility and rationality change!
0 What is rational for the swarm is not the sum of what is
rational for all individuals.
7. Prisoner’s dilemma
Prisoner B silent Prisoner B betray
Prisoner A silent A:1 day, B: 1 day A:10 year, B:0
Prisoner A betray A:0, B:10 years A:9 years, B:9 years
Two suspects are arrested. If one testifies against the other (betray) and
the other remains silent, the betrayer goes free and the silent accomplice
receives the full 10‐year sentence. If both remain silent, both prisoners
are sentenced to only 1 day for not cooperating with the investigation. If
each betrays the other, each receives a 9‐year sentence. How should the
prisoners act?
• No matter what the other player does, a player will always
gain a greater payoff by playing betray.
• Since in any situation betraying is more beneficial than
remaining silent, all rational players will betray.
15. Logistics planning
0 1991, DART (Dynamic Analysis and Replanning Tool)
scheduled transportation during the Persian Gulf
crisis.
0 Starting points, destinations, routes, conflict resolution.
0 Over 50.000 vehicles, cargo, people.
0 According to DARPA, this system alone paid back 30
years of investment in AI.