A deep-q-learning trained bot that plays Kim Kardashian: Hollywood.

ab32902641a8a08aa6392eb715c655abf0d0b2c2

1. Python environment managed with pipenv.
2. A gpu-enabled OpenCV-included installation of darknet is required.
3. Before installing python requirements you will need to install libtesseract (>=3.04) and libleptonica (>=1.71) via brew or apt-get for tesserocr to work.
4. Jython made pipenv tricky I think. I would try installing the requirements directly to the system from requirements.txt with pip3. Everything except for the monkeyrunner stuff should be running in python 3 now.

1. Name the device "vysorkim" in the window settings
2. Uncheck the "navigation bar" option in window settings.
3. You can alter video quality settings as needed.
4. Save.

• Update the MONITORS variable in config.py with a list of up to 3 monitor names and their sizes (get names from xrandr)

• TF_DEEPQ_POLICY_SAVE_DIR in config.py is the location of the saved policy to load
• TF_AI_POLICY_WEIGHTS in config.py sets the probabilistic weights of each policy in the TfAgentBlendedPolicy used in tf_ai_runner / process hub. These weights should sum to 1.0. Set any of them to 0 to remove that policy (can do all random or heuristic for testing, etc)

• sudo apt update -y
• sudo apt install python3.8
• download miniconda python 3.8 installer https://docs.conda.io/en/latest/miniconda.html#linux-installers
• Install with bash ~/Downloads/Miniconda3-latest-Linux-x86_64.sh. Select yes for last step of conda init.
• Docs here: https://docs.conda.io/projects/conda/en/latest/user-guide/getting-started.html
• Run conda config --set auto_activate_base false
• Create environment with conda create --name tf-ai-env python=3.8
• Run conda activate tf-ai-env to activate environment
• Run pip install -r tf_ai_env_pip_requirements.txt

Cuda Version 9.0.176 OpenCV Version 2.4.9.1 (from dpkg -l | grep libopencv) Might need to start redis on computer start (in the redis docs) Node.js v12.16.1 Tensorflow 1.9.0 Ubuntu 16.04 LTS 64-bit Nvidia Driver Version 384.130 Intel® Core™ i3-7100 CPU @ 3.90GHz × 4 Quadro FX NVS 810/PCIe/SSE2

Follow these instructions to prevent "login keyring did not open" error message on first chrome open: https://askubuntu.com/questions/867/how-can-i-stop-being-prompted-to-unlock-the-default-keyring-on-boot

We start the process hub on boot by using Ubuntu's built-in "startup applications" panel and running the bin/start_all_delay.sh script. gnome-terminal -e /home/cory/watson-hollywood/bin/start_all_delay.sh --geometry="180x60+0+0"

kim_current_app_monitor currently runs inside of the frontend client. maybe that could be better separated but I think its fine.

• bin - simple bash scripts to run the independent programs, used mainly in process hub
• src - all python code (tensorflow, vysor control, monkeyrunner, etc)
• apks - Android APK files for the game itself
• process-hub - Node.js program to coordinate all of the moving parts and monitor crashes, etc
• logs - Files written by process hub for each process
• frontend-static - Web javascript for rendering the frontend

1. ./process-hub/index.js
2. Thereis a command line option for --startAll false to not run the ai controller.

1. ./process-hub/run_vysor.js — Starts Vysor and opens phone window
2. ./bin/start_device_server.sh — Setup Device Server to receive commands from Device Client and send to phone (via Monkeyrunner)
3. ./bin/start_frontend_server.sh — Sets up local webserver to serve image files and state updates to web renderer
4. ./bin/start_frontend_client.sh — Opens chrome window to web renderer, and runs process to communicate with device server and ensure app stays as KK:Hollywood.
5. ./bin/start_phone_stream.sh — Start stream of images from phone to vysor for Screen Cap
6. ./bin/start_tf_ai.sh — Starts the actual loop that grabs image state from phone stream and runs actions from it