AI Roundtable (Friday) RTS/Strategy/Wargame - Josh's Notes

Randomness and Personality

• Randomization for unpredictability, and game theory solution to rock, paper, scissors
• Stars! Supernova:
  • Decision trees (producing a ranking of plans) + personality (adding a bias) + randomness (another bias)
  • 5 layer AI design using a blackboard
  • Some personalities are more random
  • Each possible strategy is rated to indicate its suitability to each personality type
  • Personality also determines inertia
• "Mixed strategy" point - decision tree should conclude that two (completely different) strategies are equally good and should be chosen 50/50; intended to avoid problem of randomness selecting between several similar strategies that are indicated by the decision tree to be near optimal
• Strategy vs. tactics - randomness is different, humans tend to have consistent strategy but varied tactics, though maybe not as appropriate for CPU players
• Kohan: composition of forces was a personality difference that was perceptable by players
• Personalities also good for game play variety, also good for reducing predictability
• User discernability important for justifying implementation cost
• Important design point: tuning predictability (sometimes predictability good, other times bad)
• If predictable, becomes a level design issue -- scripted attacks like Myth, produces more of a puzzle game
• Fog of war: some AIs play the same game as humans
• "Boltzmann" (Boltzmann soft-max) function for mapping relative utilities of actions to relative probabilities of being chosen; has a "temperature" that measures conservative vs. risk-taking
• Linking personality to diplomatic AI ends up making "reputation" useful
• Can then use reputation to stereotype human players
• The game "Diplomacy" has a very tough linguistic component (promises, ambiguity, ...)
• Kohan has an opinion/reputation matrix and personality traits (such as vindictive), would offer treaties, etc.
• From Psychology's "personality theory": Openness, conscientiousness, extroversion, apathy, neurotisism (not eroticism) ("OCEAN") were good axis to pick for describing personality. Had predictive power in experments with humans.

Diplomacy

• Nash's theorem for handling 3+ players, diplomacy
• Common goal, give & take (trade), specialization (you defend, I attack)
• Instead of common goals think mutual benefit, "coincedences of interests"
• Nash's theorm (said to) reduce the search space
• Big issue: RTSs not perfect information, so no best strategy (unlike Chess), more uncertain than standard game theory setting
• Real life: people ally with or capitulate to leader. Games: gang up on leader.
• Two separate issues: deciding who to ally with (strategy), and operational level: coordinating acts
• Operational level pretty hard right now
• Coordination similar to the porblem of arriving at the same time with units at different distances
• No vocabulary in games for communication
• Diplomacy is a game design issue since want a communication language used by both humans & cpus
• Civ II, Age of Kings, Computer Diplomacy have diplomatic actions to say "We're allied, lets go attack A".
• Kohan's sequal ("Ahriman's Gift") has a "tell" command, as does Age of Kings.
• Tell would work better if given feedback ("OK", "Later", "No")
• Age of Kings has "flares" that are drawn on strategic map, meaning "attack here" -- would be cool if CPUs participate in this dialog (it can see some spot and tell you it is important).
• Moves are communication, but thats hard to implement (in board game Diplomacy you can illegally support a unit on the other side of the map)
• Another way of communicating: give joint command over a particular battle group (USA and UK did/do this).
• Pleasing ally a goal for the AI instead of just considering action on its own merits
• Estimate the strength of other alliances in addition to the strength of one's own alliances. Feed this data into the influence map.