cpl / kalah-ai

A min-max alpha-beta pruning agent with weighted heuristics. Developed for COMP34120

Kalah-AI Gigel

Instruction

Build

In order to build our agent, you must have Go installed. The build process is simple:

cd gigel
go build

Now you should have an executable called gigel. Which you can start by simply ./gigel.

Play

The game engine for our agent is provided by The University of Manchester, for the COMP34120 course. In order to run a game between two agents use:

java -jar ManKalah.jar "<AGENT 0 STRING>" "<AGENT 1 STRING>"

Some agents use a middleman to communicate with the game engine, eg: netcat. Our agent uses localhost:12345 for the agent and localhost:12340 for a basic user "interface" that allows humans to play against our agent.

java -jar ManKalah.jar "nc localhost 12345" "nc localhost 12340"

You can simply change this inside gigel/main.go. Another two aspects of our agent that can be configured are the heuristics weights and depth of min-max tree.

Note

Please note that this is a PUBLIC DEMONSTRATION version of our agent. The development version is more advanced with analytics, better performance and overall improved code quality. The demo version is purposefully downgraded, as COMP34120 is a competitive course (and it's also fun and satisfying to find your own successful strategy).

And yes, that's why lots of variables are named hodor. It's obfuscation-ish.

Strategy

Gigel is a simple min-max search agent with alpha-beta pruning, with multiple optimizations to improve speed compared to other min-max agents. The tournament version can run at incredibly high depths, with quick decision times. Other optimizations involve low-level, compiler and multi-threading processing.

The main feature of Gigel, is the weight based heuristic. The tournament configuration was removed for this version.

Performance

With the right weights and depth, our agent can defeat all agents found in agents/ with a 100% win-rate, starting in either position, with a swift decision time of 0.5s / move.

During testing we noticed a 10-20% chance to lose and 20% chance to draw against some agents from other teams, when starting first. This can easily be fixed with more depth, increasing our win-rate to 100% and the response time to 3-5s / move.

Other agents

All agents inside agents/ are not owned, developed or in any way related to our team. They were either provided by the course staff for reference or salvaged from the tournament server. Thus our LICENSE does not extended over agents/ or ManKalah.jar.

Simulator

The kalah game simulator which can be found in gigel/simulator.go is copies the exact behaviour of the ManKalah.jar game engine, providing an interface in Go. The simulator can be extended to simulate the protocol (found in docs/).

Releases

On the GitHub repo page, you can find the tournament version of Gigel, pre-compiled for Darwin x64, Linux x64 and Linux x86. For windows users please check here. The source code is not available to the public as explained above.