journal.txt

Nov 22, 2011
combat - bug in attack where we always assume enemy will attack too
    fixed
explore - maybe instead of reverse scent diffusion, get rid of scent at all (maybe except fighter scent), and use unexplored map as desirable start this way, the frontier ants will always know where to go, and could possibly save us from having to iterate large numbers of molecular diffusion to *push* ants out as we do now. Ants deep inside empire space will just navigate based on historical value that slowly propagates into empire.
    - possible problem, this might make empire ants to prefer un-seeable space close-by(still empire space) as opposed to moving toward frontiers
strategy - when we're in strength, add desirability to un-settled regions
explore - figure out how not to lose so many ants on 1v1
    
Nov 20, 2011
combat - sometimes avoidance explore against 2 enemies will suicide
    check the code, see if we can replicate it
*strategy - when enemy hill is too close, all our ants will go for it, with noone left collecting food
    dynamically adjust enemy hill desirability
explore - ants do not seem to be very willing to help those in distress
    do some influece map plotting
strategy - keep track of blob of ants

Nov 19, 2011
strategy - keep track of spotted enemy
strategy - maybe I should really look into this symmetry stuff?

Nov 18, 2011
strategy - use multiple influence map to prioritize certain tasks (hill defense/attack, maybe something else)
strategy - depends on number of my ants and hills, use influence to keep some ants around the hill
strategy/combat - how to deal with an aggressive enemy?
    one way to deal with it is to implement that new combat logic. 
    where ants will prefer to stay at attack+1 distance when of equal strength, so it will not yield to an 
	overly aggressive opponent

Nov 15, 2011
combat - we can find all possible formations through some sort of ML learning on the spatial data, set optimal points for each *enemy* state, and only calculate at run time how to achieve a favourable state
combat - increase offensiveness when we have more friends close by

Nov 14, 2011
combat - one problem seems to be that my ant will always attack even if obstacles make standing still a better option
    - this should be remedied by 1-step look ahead idea from yesterday
exploration - when doing avoidance explore, would it be benefitial to get closer to enemy (set safe_distance to combat range) 
    so it takes advantage of more cautious enemy who tries to stay safe, but might run into trouble of unnecessary loss against a careless bot
exploration - food on the edge of viewable range is not being actively collected
    - find out exactly how diffusion spreads over multiple passes, find out the optimal diffuse factor (or even linear diffusion instead of molecular model)
combat - in winning condition, we're not pressing enemy hard enough so it can drag out to time out against a decent enemy
    - increase spread (which will push more ants into enemy's way)
        won't have immediate and easily measurable effect
        have other benefits too
        solution can be found by studying different diffusion characteristic
    - increase concentration near enemy (increase fighter ant call for help)     
        this would potentially have problem in not-winning conditions, ending up with too many fighter ants and too few exploration ants
        one remedy is to dynamically adjust this value, based on fighter ants / total ants ratio
            but still need to determine what is a good ratio
    - establish task route from my hill to enemy hill/stronghold
combat - need to distinguish situation where there are enemies on both side
    we can possibly move to one side and win
combat - regroup is still not optimal
    a distressed ant could potentially block another less distressed ant's only retreat position, while itself has other choices
    
Nov 13, 2011
combat - for each group, try to move them all in the same direction, through a voting system
    bad idea, only works better in certain edge scenarios, worse in some edge scenarios, and same in most
combat - establish a move-table, with priority and secondary preferred locatoin so higher priority ants will move to wherever they want, and lower priority ants will move to highest desirable place that's not occupied yet
        - better idea: only do this when regrouping, as formation improvement on attacking is only possible in edge cases
        - better idea part 2: use formation permutation to find the optimal formation
combat - do a 1 step look ahead, for following situations: 
    my group: press forward, or regroup behind attackradius+1
    enemy group: press forward, stay put (we can't anticipate how enemy will regroup, and for practical purpose, it makes no different)

Nov 11, 2011
1, try/resarch on if we can do gaussian blur with mask array
    no
2, check if ndimage distance can be used to do distancing in combat (or other more efficient distance checking)
    no
3, see if converting neighbour_table to numpy array, and pass it to cython function for lookup purpose will do any good
    we actually do not want too many levels of diffuse call per turn, so this is good enough

Nov 10, 2011
Idea to Revamp the influence map:
1, change it re-generate every turn, so no history is kept
	1a, this allows not setting influence on my own ant
	1b, alternatively, we can set huge negative influence on desirables to cloud over my ant's influence, without worrying about lingering effect
2, only blur from desirabilities
3, keep track of visible, explored but not visible and un-explored space, giving them appropriate influence level

ALL of the ideas above generated undesirable side effect, no go

Nov 9, 2011
I've got an idea, record the direction of each spot, and in the past n turns (say 10), where did other ants from this spot moved toward? If the majority past thresh hold x (say 50%), then we will just move toward the same direction.
This will likely cause 2 effects:
1, when starting out from a tight maze, ants will not be stuck close to ant hills bouncing around because the "observer effect" caused by other ants
2, this will cause higher chance of flocking in denser  situation, hopefully driving more ants to go toward heavy combat zones

Nov 8, 2011
Why haven't I thought of starting a journal to organize my thinking bits better? This would be way easier to navigate than random bits in that Notes.txt!

Right now, there are a few problem the AI faces:
1, not expanding fast enough, most visible in maze maps, but also slightly inefficient on open maps
2, not aggressive enough to take hills, partly related to the above, but can be solved differently
3, not protecting my own hill when there is limited resources, can be solved by solving #1, or differently
4, ant movement is not prioritized, i.e. when retreating, we want to make sure the ant in front line can retreat

basically, if we solve #1, we'd have huge improvements!