bryansweber / cunyaimodule Goto Github PK

CUNYBot, an AI that plays complete games of Starcraft.

License: MIT License

C++ 98.58% CMake 0.02% Python 1.37% C 0.03%

starcraft economics genetic-algorithm multi-armed-bandit

cunyaimodule's Introduction

CUNYAIModule

Description: CUNYBot (City University of New York) is a C++ bot that plays a full game of Starcraft:BW as the Zerg race. Its central design feature is a focus on historical economic models. In particular, it focuses on the capital/labor ratio (k), and technology/labor ratios (t), where it responds to percieved opponent's choice of k. The resulting emergent behavior can be characterized as a greedy tit-for-tat style response. Its current win rate on BASIL is about 50%.

Link to complete bot description here: https://ieeexplore.ieee.org/document/8490444

Other measures of bot performance are available at: https://www.basil-ladder.net/ranking.html#Bryan%20Weber

Recent Development Notes:

(1/23/2021) You can now play against this bot as a human on SCHNAIL: https://schnail.com/#/home

(11/17/2020) We now use grids for some portions of combat, 256x256 arrays. We also use 2 similar phases of "reservation". The first is the already existing build order system (not BOSS) executes preselected plans. It passes this plan to a short term reservation manager, so long as it does not collide with an existing reserved resource (larva, min, gas, supply). All macro actions (morph, build, research) are now reserved through this short term reservation manager, which allows for us to "multitask"- such as building lings with excess minerals while still saving for gas-heavy purchases like mutas.

(9/29/2020) Operationally, we have backed out of python and pybind11. There are various operating theaters for bots and it is challenging to find one method execution that works for all of them.

(2/19/2020) This is a huge update. The bot now uses Python via pybind11. This permits (easy) access to the python libraries, but greatly increases the size of the bot. Installation now requires manually downloading python36.dll and using pip to install the requisite dependancies. An option to compile without these dependancies is still available.

(9/26/2018) Incorperated FAP fully, BWEM is now used for pathing and movement, and BWEB is now used for building, including some walls at the natural. Several functions have been deeply refactored (particularly workers) and combat has been deeply streamlined for student use. The bot now makes some estimates of what opponents may have behind the fog of war, and better responds to observed technical investments.

(9/6/2018) Current drafts of CUNYBot are responding to opponent's choice of t as well, and the overall observed behavior remains a tit-for-tat style. Technology has been closely integrated and better improved in current versions, but have not yet released these to SSCAIT. Tasks at hand are generally production choices, and reducing waste in combat. To that end, FAP has been added as a combat simulator. It also serves as an agnostic method of choosing which units to produce, with some caveats.

cunyaimodule's People

Contributors

Stargazers

Watchers

Forkers

shamigiri johnwjackson kikwon

cunyaimodule's Issues

Checking if minerals are safe. Reapply simple checks.

Expo locations are sometimes missed/inaccurate

Complete self-wall-offs are possible.

Potentially prevent building directly on map veins 150+?

Freeze on lair upgrade.

Travis Shelton blocks expansions.

Project now responds to opponent's choice of Army vs Worker building.

It matches their visible units, and assumes continuous building of workers off of one base to a maximum of 85. The true rate is (likely) much quicker than that.

Drones die in transfer. In large quantities.

High Ground is 47% miss chance, in addition to concealment.

might want to program that in somehow.

53.125% hit chance, 1- for miss chance.

Reorganize building queue system, get stuck on expos often.

Upgrade section hits 90 ms ping? How?

Expos only to visible expansions.

Building orders check for "active builders", this is likely creating some freezes.

Worker defense?

Defeat UAB's opening probe rush please.

Scourge shouldn't suicide into interceptors.

Check for enemy unit value first.

Bot has weird tendency to upper corners?

Potentially a 0.0 bug? See game with Hannes.
http://www.openbw.com/replay-viewer/?rep=https%3A%2F%2Fsscaitournament.com%2FReplays%2FBRYAN+WEBER%2F248409-Brya_Hann-ZvT.rep

Units may conga line into enemy base.

The case appears to be one where:

The closest e unit (by air) is a weak or isolated combat unit.
The ground path to the e unit is convoluted and blocked by a superior e force.
Friendly unit then is "suprised" by enemy force along path.

Bot sometimes does 9-hatch?

Vs P, La Mancha, Top Right Spawn.

gas carring worker to build expo derps out

Opening learning should use Bayes' Theorem.

I mean, why wouldn't you use it?

Missing Minerals

Misses an entire mineral patch on destination. Seems to happen on a few other maps.

Static Defence is wildly undervalued.

What do we do about it?

Actual measure of distance. Fill web out and in.

Drones are being double-counted by minerals.

Problem: one drone is assigned to a mineral.
Mineral increments.
Drone is assigned elsewhere (?)
Mineral is not decremented.

Solution outline: Have each mineral track the names of the units attached to it every frame?

Expansion check should be done if worker is moving as well.

double extractor is annoying and wastes a worker.

Attempted fix deployed.

have bot choose from multiple build orders

Units are blind to nearby enemy ranged units.

Missed natural

Benzene? (3 player map left side spawn

workers sometimes fail to build, likely spamming drones.

Units don't recognize they're being killed as they fight.

E-local eminates from the target.
Friendly-local eminates from the self.

There is a non-overlapping area (especially noticeable for long long range units), where an enemy unit can be OUTSIDE of range of my target, but able to attack me. (usually from the side or behind or something)

Economy ramping is unusual around 2 hatch period.

Too much army investment.
Too much idle tech investment.

Returns to normal somewhere around 3-5 hatches.

Bot occasionally loses track of inventory.

Time of problem: Often mid-game, close to 140ish control.
Perhaps this is related to the death of minerals that usually occurs at that time.

Description of problem: I can see rapid fluctuation in the count of workers, and larva and eggs.
Sometimes, I see variation in the counts of hydralisks as well.

Consequence of problem: The bot then is unable to establish its state and freezes. But not all units. Perhaps only those on the field at the time of the crash? The bot often becomes unstuck after massive losses.