The project aims on building a new type of Artificial intelligence algorithm which is simple and surpasses many already available algorithms for Humanoid or Mu-Jo-Co(Multidimensionla-Joint-with-contact) locomotion related tasks. It simulates a powerful AI Algorithm,called Augmented Random Search (ARS) by training a Half-cheetah (Mu-Jo-Co) to walk and run across a field. to walk and run .
Link to the Google-DeepMind's Video
- Asynchronous Actor-Critic Agents
- Deep Learning
- Deep Reinforcement Learning
- Unlike other AI systems where the exploration occurs after each action (Action Space) , here exploration occurs after end of each episode (Policy space)
- ARS is a shallow learning technique unlike deep learning in other AI's systems (Uses only one perceptron rather than layers of it)
- ARS discards the technique of Gradient Descent for weight adjustment and uses the Method of Finite Differences
- Perceptrons
- Reward Mechanism and updation of weights
- Method of finite Differences to find the best possible direction of movement
- Scaling the update step by standard deviation of Rewards.
- Online normalization of weights.
- Choosing better directions for faster learning.
- Discarding directions that yield lowest rewards.
- Fork and clone the repository using
git clone https://github.com/ashutoshtiwari13/Simple-Random-Search.git - Run
pip install -r requirements.txt - Also check the Simulation.txt for setting up the PyBullet Simulation Environment
- Use the Anaconda Cloud - Spyder IDE (Any framework/IDE of your choice)
- Use Python 3.6 and above
- Run the command
python ars.py
Rewards start from being negative as low as -900 and climbs to positive 900 in around 1000 steps.
Happy coding 😊 ❤️ ✔️