https://github.com/threewisemonkeys-as/rlosfest_2022_screening_exercise
https://github.com/threewisemonkeys-as/rlosfest_2022_screening_exercise
Last synced: 2 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/threewisemonkeys-as/rlosfest_2022_screening_exercise
- Owner: threewisemonkeys-as
- Created: 2022-04-05T02:41:59.000Z (about 3 years ago)
- Default Branch: main
- Last Pushed: 2022-04-05T03:03:12.000Z (about 3 years ago)
- Last Synced: 2025-02-09T20:43:32.221Z (4 months ago)
- Size: 1.95 KB
- Stars: 0
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# RLOSFest 2022 Screening Exercise
## What compiler optimization tasks do you want to tackle and why?
As part of the project, I want to build agents to tackle the [`LLVM`](https://compilergym.com/llvm/index.html) optimisation environment in `CompilerGym`.
This environment essentially encapsulates a sequential decision making problem in which the agent has to generate a sequence of optimisation transforms for a given input program that optimises for some objective.
Useful objectives could include
- Minimising IR Instruction Count (cheap to evaluate)
- Minimising Memory used during runtime (expensive to evaluate)
- Minimising runtime (expensive to evaluate)
- A learned metric that takes as input some program representation and outputs estimated runtime (cheap to evaluate during test time)
The main reasons for tackling this particular task are
- It can be naturally modelled as a sequential decision making problem which allows us to use a large range of tools from _search_ to _reinforcement learning_
- Preliminary results from the `CompilerGym` leaderboard show that significant gains over the default `-O3` flag are possible even using random search (e.g. 1.5x for `sha`). This shows promise for a learning based method which understands input program semantics to give significant gains over default confguration. Such a tool could have a massive impact even if it gives consisitent fractional improvemnent since LLVM is used across the industry.
## Links to past coding project in Python or C++
### Open Source -
1. [GenRL](https://github.com/SforAiDl/genrl): PyTorch Reinforcement Learning Library
- Contributed implementations of various Deep Contextual Bandits
- Core Maintainer and worked on implementation of [distributed RL](https://github.com/threewisemonkeys-as/genrl/tree/distributed/genrl) using RPC
2. [Trotbot](https://github.com/ERC-BPGC/Trotbot): Autonomous Delivery Robot
- Built obstacle detection and path planning stack using Robot Operating System (ROS) in Python
- Implemented Rapidly Exploring Random Trees (RRT) for path planning in complex indoor environments
3. [GenNav](https://github.com/ERC-BPGC/gennav): Python library for Robotics Navigation
- Co-author and Lead Maintainer working with a team of 10+ student contributors
- Modular collection of navigation algorithms and utilities commonly used in Robotics with a ROS wrapper
### Personal Projects (also open source) -
1. [Causal Reasoning from Meta-Reinforcement Learning Exploration](https://github.com/threewisemonkeys-as/causal-meta-rl/)
- Devised, performed and documented additional experiments to futher evaluate the central claim that Meta RL
agents can performs Causal Inference
2. [Torched Impala](https://github.com/threewisemonkeys-as/torched_impala): Attempt at implemented distributed deep RL in PyTorch
3. [Drone Controller Wrapper using MAVROS and PX4](https://github.com/threewisemonkeys-as/drone_automation)