This is a demo where FCL fights against an Intel agent from the vizdoom competition.
apt install docker.io
systemctl start docker
systemctl enable docker
Edit /etc/group and add your username to the group "docker".
Build these two dockers:
./build.sh intelact
./build.sh host
Clone vizdoom with:
git clone https://github.com/mwydmuch/ViZDoom.git
Install it with:
sudo pip3 install .
First run the host, followed by the bots:
./run.sh host
./run.sh intelact
The best approach is to run this in two separate terminal windows.
Our own FCL agent runs just in the Linux system itself and connects to the host.
To run the DFL bot:
python3 ./run_agentFCL.py 0.0001
where 0.0001 is the learning rate.
Which runs a deathmatch for the length of the time specified
in host/host.py, and will write to 3 files:
KD.txt
wtDist.txt
FCLOutput.txt
The first logs a timestamp each time the bot is killed.
To calculate a KD ratio, you will need to edit
intelact/IntealAct_track2/run_agent.py to write
the corresponding death events for the intel bot.
wtDist.txt logs the average Euclidean distance
of the weights from their initial values, per layer.
FCLOutput.txt logs various measures:
feedback error, bot steering action, health. See the code.
There are a number of *.cfg files here - if you want to run deathmatch in the way described above, copy _vizdoomNetwork.cfg to _vizdoom.cfg. Assuming the host is already running, running run_agentDFL.py will then attempt to connect and start a match. It's crucial that all bots, and the host, are using the same Doom map. This is specified in their own .cfg files.
You can use any of the .wad files that come with VizDoom, but we recommend editing them to a) have more spawn points so that the bot gets a better randomisation of starting position and increase generalisation of learning, and b) to increase the amount of ammo available, otherwise the enemy can run out of ammo and start behaving strangely. We used http://www.doombuilder.com/ for this.
If you want to change the actions available to the bot, you need to edit config/config.cfg. At the moment, the bot only rotates in the plane, and shoots, so there is lots of scope for improvement!
The intelact robot uses the CPU for its machine learning which in turn uses tensorflow. Feel free to install the tensorflow library which GPU support instead.