Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/cbeck88/wesnoth_mclp_ai

An AI project for wesnoth
https://github.com/cbeck88/wesnoth_mclp_ai

Last synced: 2 months ago
JSON representation

An AI project for wesnoth

Awesome Lists containing this project

README 

        
Wesnoth_MCLP_AI

===============



An AI project for wesnoth. 



The idea is that the AI should evaluate the quality of a hypothetical position by running monte carlo simulations. 

During these monte carlo simulations, all parties will be controlled by an efficient and aggressive AI, which seeks

to kill as much gold as possible in its turn. To do this, it solves a series of linear programs related to the game

state, hopefully very quickly. The linear programs seek to approximately determine



1. Which enemy unit is easiest for me to kill?



2. What is the opportunity cost of attacking with a particular unit?



The LP_ai heuristic is to *attack the unit which can be most easily killed using all adjacent slots, using the 

unit from the optimal solution which is otherwise the least useful first*. After we see if we got our hits, 

reevaluate the same way.



The MCLP_ai scores moves on the basis of monte carlo experiments with LP_ai as the controller.



Current AIs

-----------



**lp_1_ai**: This solves an LP for question (2) above. It assigning all units to slots to attack all enemies

in the way which maximizes expected damage. The ai executes these moves to display the optimum solution.

By removing a unit and a slot and recomputing the OPT, you can estimate the opportunity cost of an attack.

(might be a good idea to make a gold-adjusted version of this as well)



**lp_2_ai**: This solves an LP for question (1) above. For each enemy unit, list all possible ways it can be 

attacked and from which slot, then find attack solution which maximizes ctk. We approximate ctk with a fractional LP. 

The ai solves this for every enemy unit, picks the best one and displays the optimal solution by moving 

these units to attack.



Both of these are just tools to visualize the results of these LPs for further testing. 



**lp_ai**: A helper ai using the lps tested in lp_ai_1, lp_ai_2, which seeks to execute turn as aggressively 

as possible, using maximum slot efficiency. In a preliminary state.



**mclp_ai**: Uses lp_ai in monte carlo simulations. In a preliminary state.



*These ais will not recruit or grab villages.* 

You should load them from a savegame file which is already set up for an attack. This is most easily done using the tools in testsaves/. 



mclp_ai is the only one which is planned to become a full-fledged ai which recruits and grabs villages.



Other

-----



There are some test saves and utilities in /testsaves. Use the script aitest to inject an ai into one of the savegames

found there, and load it into wesnoth-1.11. Or add your own replays/ais for testing.



I may decide to write lua hooks to the Monte Carlo procedures, so that this idea can be reused in the future more easily.



LP's are solved using the lp_solve library 

(API, 

example). 

I installed from a linux mint package *liblpsolve55-dev*, 

following instructions here.



You must have this library to build.



*If you are having trouble linking it:* For me the package automatically put a library file liblpsolve55.a in /usr/lib/, and according to scons output, g++ is told to look for libraries there, 

so I didn't have to do much. I don't know anything about scons though so if it doesn't work out of the box good luck :)

You also need libcolamd2.7.1 package at least. For me this was automatically there. You need glibc (= library ld?) 

version 2.2.5 (if I remember the page correctly). This version is very old so almost surely whatever you have is fine.



Building

--------



I assume that you already have the ability to compile wesnoth from source. If not you can visit our fine friends over 

here.



MCLP AI is essentially a branch of wesnoth-old. I just give it its own github repository like this to keep it at a 

manageable size.



To build MCLP AI, (lasted tested at wesnoth commit: 5a0952a9fe6703623d4f528ca09)



1. Make a folder to hold both projects, e.g. ~/wesnoth-src. Clone wesnoth-old and Wesnoth\_MCLP\_AI into this folder.



   (If MCLP_AI base folder is not a sibling of wesnoth-old, the only bad thing is that the testsaves suite won't work.)



2. Now the contents (SConstruct, /data/, /src/) of MCLP AI are intended to be installed (copied) to the base folder wesnoth-old.



   You might want to "git checkout -b MCLP_AI" in wesnoth-old beforehand.



3. Then just run scons in /wesnoth-old/ as usual.



The ~~only~~ core files which are modified are /src/ai/registry.cpp, /src/SConscript, and /SConstruct.



Because of what I consider to be bugs in AI module, 

it is not possible in the current version to add a custom C++ ai without altering several additional core AI module files:



/src/ai/manager.cpp, /src/ai/composite/ai.hpp



Hopefully a future patch will make this unnecessary.



In _/wesnoth-old/src/ai/registry.cpp_: Add my ai factories so that they can be used in game.



    diff --git a/src/ai/registry.cpp b/src/ai/registry.cpp

    index cc57dcf..0435fc4 100644

    --- a/src/ai/registry.cpp

    +++ b/src/ai/registry.cpp

    @@ -48,7 +48,8 @@ static register_ai_factory ai_factory_default("");

     static register_ai_factory ai_default_ai_factory("default_ai");

     static register_ai_factory ai_idle_ai_factory("idle_ai");

     static register_ai_factory ai_composite_ai_factory("composite_ai");

    +static register_ai_factory ai_lp_1_ai_factory("lp_1_ai");

    +static register_ai_factory ai_lp_2_ai_factory("lp_2_ai");

     

     // =======================================================================

     // Engines



In _/wesnoth-old/src/SConscript_: Add my source .cpp files



    diff --git a/src/SConscript b/src/SConscript

    index fbbe281..e6bc25d 100644

    --- a/src/SConscript

    +++ b/src/SConscript

    @@ -197,6 +197,7 @@ wesnoth_sources = Split("""

         ai/game_info.cpp

         ai/gamestate_observer.cpp

         ai/interface.cpp

    +    ai/lp/ai.cpp

         ai/lua/core.cpp

         ai/lua/lua_object.cpp

         ai/lua/unit_advancements_aspect.cpp

         

         

to the list of wesnoth sources, with the other ai.cpp files, to get scons to build it. (note: this is out of date, there are more cpp files now)



In _/wesnoth-old/SConstruct_: Add my required libraries so that they are checked and statically linked.



    diff --git a/SConstruct b/SConstruct

    index 5d4749c..722a158 100755

    --- a/SConstruct

    +++ b/SConstruct

    @@ -338,7 +338,11 @@ if env["prereqs"]:

             conf.CheckSDL("SDL_mixer", require_version = '1.2.0') and \

             conf.CheckLib("vorbisfile") and \

             conf.CheckSDL("SDL_image", require_version = '1.2.0') and \

    +        conf.CheckLib("liblpsolve55") and \

    +        conf.CheckLib("colamd") and \

    +        conf.CheckLib("dl") and \

             conf.CheckOgg() or Warning("Client prerequisites are not met. wesnoth, cutter and exploder cannot be built.")



I am not going to fuss with required versions in the SConstruct script,

because scons is something of a mystery to me and I don't want to mess around with it.



In _/src/ai/composite/ai.hpp_



    diff --git a/src/ai/composite/ai.hpp b/src/ai/composite/ai.hpp

    index 7d0685c..9ef61d1 100644

    --- a/src/ai/composite/ai.hpp

    +++ b/src/ai/composite/ai.hpp

    @@ -52,7 +52,7 @@ public:

            /**

             * Play the turn

             */

    -       void play_turn();

    +       virtual void play_turn();

     

     

            /**

    @@ -66,7 +66,7 @@ public:

            virtual void new_turn();

     

     

    -       std::string describe_self() const;

    +       virtual std::string describe_self() const;

     

            /**

             * serialize

    @@ -95,7 +95,7 @@ public:

            void create_engine(std::vector &engines, const config &cfg);

     

     

    -       void on_create();

    +       virtual void on_create();

     

            /**

             * unwrap



In _/src/ai/manager.cpp_



    diff --git a/src/ai/manager.cpp b/src/ai/manager.cpp

    index dea1dc9..6c36445 100644

    --- a/src/ai/manager.cpp

    +++ b/src/ai/manager.cpp

    @@ -18,6 +18,7 @@

      */

     

     #include "composite/ai.hpp"

    +#include "lp/ai.hpp"

     #include "configuration.hpp"

     #include "contexts.hpp"

     #include "default/ai.hpp"

    @@ -81,7 +82,16 @@ void holder::init( side_number side )

                    default_ai_context_ = new default_ai_context_impl(*readwrite_context_,cfg_);

            }

            if (!this->ai_){

    -               ai_ = boost::shared_ptr(new ai_composite(*default_ai_context_,cfg_));

    +                //this case analysis added by iceiceice nov 23 2013, as the quickest hack that would work

    +                if (cfg_["ai_algorithm"] == "lp_1_ai") {

    +                    ai_ = boost::shared_ptr(new lp_1_ai(*default_ai_context_, cfg_));

    +                }

    +                else if (cfg_["ai_algorithm"] == "lp_2_ai") {

    +                    ai_ = boost::shared_ptr(new lp_2_ai(*default_ai_context_, cfg_));

    +                }

    +               else {

    +                    ai_ = boost::shared_ptr(new ai_composite(*default_ai_context_,cfg_));

    +                }

            }

     

            if (this->ai_) {