TDP-MSRC-AI-Challenge

This project is created for The Malmo Collaborative AI Challenge, which concerns writing an AI-agent for a game called Pig Chase created in Minecraft. Turn to https://www.microsoft.com/en-us/research/academic-program/collaborative-ai-challenge/ for a description about the game and the challenge.

Running the code

1. Install the challenge code as described under installation here
2. Go into the ai_challenge folder in the malmo-challenge and run

wget https://github.com/philipjhj/TDP-MSRC-AI-Challenge/archive/master.zip
mkdir danish_puppeteers
unzip master.zip -d danish_puppeteers
mv danish_puppeteers/master/* danish_puppeteers/
rm -r danish_puppeteers/master/
cd danish_puppeteers

Add the pig_chase and danish_puppeteers folders to your python path by using the command export PYTHONPATH=$PYTHONPATH:<path to ai_challenge>/ai_challenge/danish_puppeteers respectively. After this you should be able to run the evaluation script python pig_chase_eval_sample.py or any of the other scripts.

To run a script with docker on an Azure machine created with docker-machine, run

./run_azure_docker.sh <machine-name> <python-script-name-without-file-extension>

See more about how to run the challenge with docker here

Idea

Our agent makes a decision of whether a challenging agent is cooperative or not. After making this decision all following actions are computed using traditional AI-techniques (A-star search in game-graph using the available actions etc.). The determining factor is thus whether the agent can correctly identify the intentions of a challenging agent. We have made two different modes of how it does this.

Mode 1: Guided Danish Puppet

The first mode uses a simple heuristic to determine whether the challenging agent is cooperative. If the challenger moves towards the pig with 60% of its moves, then it is assumed cooperative otherwise it is considered non-cooperative. This solution is very specific to the target game and does not include any machine learning methods. It is basically an improved version of the A-star agent, in which A-star is run on multiple targets

Mode 2: Stringless Danish Puppet (AKA Pinocchio)

The other mode uses a Hidden Markov Model in an attempt to model the intention of the challenger. It uses a few observable variables as input:

1. The helmet color
2. Sign of difference in distance to the pig (takes values -1, 0 or 1)
3. Sign of difference in distance to nearest exit (takes values -1, 0 or 1)

It then computes the marginal distribution over the usefulness of the following action and uses this probability to determine whether the challengers intentions are cooperative or non-cooperative.

More information can be found in the Documentation-file.

Demo

See a video here demonstrating our agent in action.

Results

Our results based on the pig_chase_eval_sample.py script can be seen here. The experiment name matches the method used.

• We had technical issues with the evaluation script when we ran it with our agent using the HMM.

Post-competition results and work (especially with respect to the above mentioned problem) may be found in the Developement-branch.