Principle

Organize lessons from diverse strategical sources encountered including games, books, or any kind of activities that lead me to improve strategical thinking. Also explore new possible solutions. Note that strategy is solely a mean to reach a goal, for example applying EthicalFramework and improving redistribution to a fairer share.

Lessons ordered by importance

  1. having a goal
  2. having a strategy to reach it

Games played

  • Chess
    • fork, patterns, time pressure, non-key moves, meta-strategy, opponent discovering, long-time learning, ...
  • Go, cf MentalExercises#Go
    • division for zone control, connectivity, fork (2 Atari), patterns (eyes, life and death)
  • Quake III
    • timing, alocentrism, compound advantage, position control, strategical theory of mind
  • Warcraft III
    • micro-management and macro-management differences

Books

new games / AvancedRTS

Starcraft 2

Why am I watching SC2 replay at all?

  • expecting to find more general patterns that I can apply to other domain

On-going questions

  • is there a hierarchy of strategies matching the player ranking, i.e. waiting for new builds by the top players and seeing them trickle down the hierarchy
  • can effective strategies be summed up via StructuralInformationAsymmetries?
    • establishment and maintaining of compounding over time cheap to exert and hard to counter pressures
    • to constantly pick amongst the set of possible actions the one that will at the same time satisfy
      1. maximizing my advantage
      2. minimizing my opponent advantage (distinguished from #1 because I have less control, I can only influence)
      3. minimize the amount of information the opponent have about the path of action (to handle the meta-game and learning across time)
      4. make this decision making increasingly cheaper (to take into account computational resources available, note that this is not limited to the current "session" but rather to other sessions following the same rules and even other sessions with different rules and enforce re-organization optimization)
    • to make the tipping of the balance (creating power asymmetry) for myself easier and countering harder for the opponent
      • rather than to tip the balance by applying the same amount of work or worst an increasing amount of work
    • Discussion:blinkenshell/trunkie.log (15/07/2011 4:50PM) and Discussion:freenode/#gameai.log (15/07/2011 5:50PM)
  • is this traversing Network Of Strategies heuristic valid?
    1. on the your situation (map + race + position) have an optimal build?
    2. if yes, apply it
    3. if not, does the opponent have one?
      1. if yes, apply the best counter
      2. if not, hypothesize on your opponent build (based on statistics of his previous plays) and start the best counter
        1. gather information to validate or not your hypothesis
        2. repeat
  • what is the impact of new indicators?
    • e.g. army size by cost, number of harvesters
      • plenty of economy-based comments e.g. considering the ROI of a unit or dedicating a lot of APM for an action (in the sense that actions themselves are a resource and thus their allocation matters)
  • can we sum-up the economy as income curves under which productions is allowed?
    • thus having a progressive increase until saturation then plateauing until the next expansion is done
      • with temporary decreases during harassment
    • minimal threshold under which it becomes impossible to bootstrap again
      • 1 town hall + 0 worker + income < cost of worker
      • 0 town hall + 1 worker + income < cost of town hall
    • the curve could also take into account infrastructure (or administrative) costs
      • thus pushing the relative curve down
      • base + drones + supply depots
      • not considering long distance mining
    • periodic timer to notify for required expansion
      • based on map and depletion time
      • e.g. expand every ~3min (modulo -threat +push)
  • fitness function
    • always modulo time since it is with an opponent
      • army size
      • total damage potential
      • " " " per category (ground, air)
      • " " " " " multiplied by the inverse cost of available counters
    • also note that the value of the destruction of any item decrease over time
      • removing an enemy probe very early in the game is much more important than later on
  • the ability to evaluate forces and positioning decides engagement

Network Of Strategies

  • own heuristic
    • it is equivalent to routing over a network of builds linked to their counters
      • refined builds can initially be clustered to limit complexity at the cost of precision
        • first modeling can be limited to popular ones
          • within a restricted community
            • because of patches in particular, starting with a competition could be much more efficient than accumulating data
          • the more general step could use a replay analysis tool, as tried before, based on BO
            • the underlying hypothesis is that it is possible to backtrack from produced units (target) to build order (means) to strategies
      • listing Starcraft 2 Counters List as typed counter links (edges) between strategies (nodes) (or Counters by Unit since one can consider the outcome of a strategy to be mainly the set of units produced)
        • ideally this would not be based on curated list but rather directly from unit data after each patch, equivalent to an inferred trophic web
      • initial situation of both parties highly benefit from prior information, including historical data
      • links weight do change based on updates, especially economical value
      • add typed transition links between nodes from the same race and build order with one action
        • link weight would display the cost
        • this might also need clustering to maintain readability
    • this would benefit from applying the same visualization suggested by LeveragingRandomness and used in FabelierGephiWorkshop
    • existing wiki (e.g. Team Liquid) could be exploited to map strategy per race then counter links
      • AWS#MTurk could be used to gather training data on strategy use per player per period
  • mirror strategies only work when the opponent is already using the dominant strategy and with the same race
  • transition between strategies should be listed
    • increasing complexity but diminishing overlap and allowing to be more exhaustive
    • possibility to tag strategies by game moment, e.g. opening/mid-game/end-game
  • the main opportunity might be in edges that become weaker after a patch
    • strategies that were until then useful and used very often but have significantly lose efficiently yet are used out of habit
    • (HD258) oGsEnsnare vs CreatorPrime - PvT, PomfEtThud May 2011 discusses the possible new strategies before a patch get released
  • the network could be displayed in real-time and each information gathered would prune it

Example

Nodes and Edges should be saved as .csv files then loaded into FabelierGephiWorkshop

Nodes
Id,StrategyName,Race,MinMinerals,MinGas,MinTime,TimesUsed,DateFirstUse,BuildOrder,Details
1,MMM,Terran,,,,,,
2,RoachBust,Zerg,,,,,,
Edges

Notes that edge are unidirectional, going from the source to the target representing the dominating strategy supposing everything is conducted perfectly by both players (no fluck nor specific advantage from the map).

Source,Target,Details
1,2,

What am I actually looking while watching for a replay?

  • new creative strategies
  • example of build orders
    • e.g. 1-1-1
  • micro-management tips
    • e.g. units positioning, fortifications
  • macro-management views
    • efficiency of a build against another
    • when and where to expand
  • keep track of updates
    • e.g. deprecated builds
  • what are the crucial decisions and when are they taken?
    • build order?
    • expand or not?
    • attack or not?
    • retreat or not?
    • switch unit production?

Extracted generic strategy heuristics

  • micro for a macro goal
  • have and apply default macro
    • based on
      • patches
      • opponent specie and latest games
      • map
    • gather info and adapt production against opponent macro
      • have a hierarchy of macros and their counters
  • micro
    • constantly focus against on natural counter with lowest life points
    • harassment has to be at least but ideally both cognitively and cost effective
    • consider during an encounter the cost of units
      • i.e. do not lose costly units to cheap units even if the battle can be won (except if it leads to another advantage, e.g. stopping an expansion)

To find

  • cost over time of units/upgrades

Super Cluedo

Based on our previous game here are few ideas that I think could be useful, sorted by importance :

  • focus on one category (weapon, person or location) by making an hypothesis that contains one or two of our own card
  • write down the first letter of who shows a card as it allows to see who has shown all their cards, track who still has cards to show
  • when nobody shown anything the category is either the right one or held by one of the NPC
  • use Search at the beginning of the game then focus on hypothesis
  • track who does what mistake to known who is closer and hence if it makes sense to take a risk (e.g. making an accusation while still having 2 possibilities in 1 category)
  • listen to who is making what hypothesis or avoiding specific location in order to influence your own hypothesis or accusation
  • use an order (e.g. alphabetical) to go systematically through the items of each category

Sources

Inspired by

See also

To do

  1. link to
    1. WikBrain mapping (neo-cortex?)
  2. finance
    1. application to ATS
  3. my own presentation of Strategy vs Tactic
    1. IUT Lannion 2002
  4. General Game Playing (GGP)
  5. Skynet meets the Swarm: how the Berkeley Overmind won the 2010 StarCraft AI competition by Haomiao Huang, arstechnica January 2011
  6. apply NetworkOfStrategies to other fields
    1. e.g. business competition in a specific market (inspired by TenDayMBA#Chapter9) but replacing
      1. game rules by the legal framework
      2. unit values by solutions
      3. build order by production costs
      • MainOpportunity thus translates here too by shifts except here they can be intrinsic or at least heavily biased by previous investment
  7. consider a short-term equivalent of CognitiveEnvironments#TasksByEfficiency for any real-time strategy game
    1. eventually with different patterns for each stage of the game (opening/mid-game/end-game)
    2. e.g. have a timer for expansions, etc...
  8. http://www.quora.com/What-are-some-lessons-learned-through-playing-StarCraft-that-are-useful-in-real-life