Skip to content

Latest commit

 

History

History
executable file
·
92 lines (64 loc) · 7.05 KB

README_agents.md

File metadata and controls

executable file
·
92 lines (64 loc) · 7.05 KB
Raw

Intelligent Agents in Komondor

Table of Contents

Authors

Both authors have the same contribution to this project.

Introduction

Komondor includes integrated Machine Learning (ML) operation, which is implemented based on the ITU-T Architecture for Future Networks. In particular, several decentralized and centralized ML mechanisms have been provided to fulfill use cases such as Decentralized Learning-based Spatial Reuse (SR) or Smart Dynamic Channel Bonding (DCB).

For more details on the ML-aware architecture, please refer to Wilhelmi, Francesc, et al. "A Flexible Machine Learning-Aware Architecture for Future WLANs." arXiv preprint arXiv:1910.03510 (2019).

Overview

The modules related to agents operation are as follows:

  • central_controller.h: entity that is connected to agents that indicate support for the centralized operation. Allows centralizing procedures such as data gathering or ML output generation.
  • agent.h: entity that is connected to Basic Service Sets (BSSs) through the Access Point (AP). The agents collect data from BSSs and can provide new configurations to them, based on certain optimization mechanism.
  • learning_modules: here we find the implementation of ML methods that receive feedback about the networks performance in simulation time.
  • agent_methods.h in methods.h : auxiliary methods used by agents.

An overview of the agents operation is next shown:

Usage

Detailed installation and execution instructions can be found in the Komondor User's Guide.

To run Komondor with intelligent agents, additional information should be provided to the simulator. In particular, you should run Komondor simulator by using the following command:

$ ./komondor_main INPUT_FILE_NODES INPUT_FILE_AGENTS OUTPUT_FILE_LOGS SIM_CODE FLAG_SAVE_NODE_LOGS FLAG_SAVE_AGENT_LOGS FLAG_PRINT_SYSTEM_LOGS FLAG_PRINT_NODE_LOGS FLAG_PRINT_AGENT_LOGS SIM_TIME SEED

The inputs are further described next:

  • INPUT_FILE_NODES: file containing nodes information (e.g., position, channels allowed, etc.).The file must be a .csv with semicolons as separators.
  • INPUT_FILE_AGENTS: file containing agents information (e.g., ID of the BSS to which the agent is associated, ML mechanism, inter-request interval, etc.).The file must be a .csv with semicolons as separators.
  • OUTPUT_FILE_LOGS: path to the output file to which write results at the end of the execution (if the file does not exist, the system will create it).
  • FLAG_SAVE_NODE_LOGS: flag to indicate whether to save the nodes logs into separate files (1) or not (0). If this flag is activated, one file per node will be created.
  • FLAG_SAVE_AGENT_LOGS: flag to indicate whether to save the agents logs into separate files (1) or not (0). If this flag is activated, one file per node will be created.
  • FLAG_PRINT_SYSTEM_LOGS: flag to indicate whether to print the system logs (1) or not (0).
  • FLAG_PRINT_NODE_LOGS: flag to indicate whether to print the nodes logs (1) or not (0).
  • FLAG_PRINT_AGENT_LOGS: flag to indicate whether to print the agents logs (1) or not (0).
  • SIM_TIME: simulation time
  • SEED: random seed the user wishes to use

IMPORTANT NOTE (!): Setting FLAG_SAVE_NODE_LOGS and FLAG_SAVE_AGENT_LOGS to TRUE (1) entails a larger execution time.

Agents Input Files

The agents' operation is ruled by the agents input file. The current parameters are as follows:

  1. BSS code: a code that must match with BSS codes provided in node input files. "NULL" indicates that the agent acts as a central controller.
  2. Centralized flag: flag that agents use to indicate their support to the centralized system.
  3. Time between requests: an agent is supposed to request data to the AP every inputted period. For the central controller, the time between requests is meant for requests to agents.
  4. Allowed actions as lists: for each type of modifiable parameter, the user must introduce a list of possible values separated by a comma (e.g., CCA = {-70, -75, -80, -82}). Currently, there is support for lists of channel, CCA, transmission power, and DCB policy.
  5. Reward: type of reward to be used by the learning mechanism (e.g., REWARD_TYPE_PACKETS_SUCCESSFUL = 0, REWARD_TYPE_AVERAGE_THROUGHPUT = 1)
  6. Learning mechanism: learning framework used by the agent/CC (e.g., MULTI_ARMED_BANDITS = 1)
  7. Selected strategy: extra parameter to differentiate between subtypes of learning approaches. For instance, in the bandits framework, we have different action-selection strategies (e.g., STRATEGY_EGREEDY = 1).

See an example of agents input file here.

Regarding the output ("output" folder), some logs and statistics are created at the end of the execution.

An example with agents: finding the best primary channel

Assume that an AP implementing dynamic channel bonding wants to find the best primary channel defined as the primary resulting in the highest throughput. Consider the deployment below, where all the BSS's implement single-channel transmissions but AP A, which is allowed to bond channels. Assume also a fully-overlapping setup, where every node listens any transmission falling within its primary channel, and where all APs are fully backlogged. In such a situation, what would be the best primary channel for AP A?

Well, since channels 4 and 5 are always free, thus providing the largest interference-free bondable combination (40 MHz) permitted by IEEE 802.11ac/ax, A should select either 4 or 5 as primary. How would A learn that? In this example, we propose a simple combination of the MAB mechanism together with the epsilon-greedy learning strategy. In the charts below, we observe that the agent successfully tends to pick primary 5 since it results in the highest reward (defined as normalized throughput). Notice, that 5 is selected rather than 4 due to pure randomness in the algorithm initialization.

You may find the input files for the nodes and the agent of the example here.

GO BACK TO MAIN PAGE