pkyriakou / RL-reward-experiments

This repository is a fork of a repository originally created by Lucas Descause. It is the codebase used for my Master's dissertation "Reinforcement Learning with Function Approximation in Continuing Tasks: Discounted Return or Average Reward?" which was also an extension of Luca's work.

RL Reward Experiments

For documentation regarding the tasks used in these experiments refer to the following repository: RL-Continuing-Tasks

This is a fork of Lucas Descause's original repository. It is the codebase used for the experiments of my master's dissartation "Reinforcement Learning with Function Approximation in Continuing Tasks: Discounted Return or Average Reward?"

Abstract

Reinforcement learning is a machine learning sub-field, involving an agent performing sequential decision making and learning through trial and error inside a predefined environment. An important design decision for a reinforcement learning algorithm is the return formulation, which formulates the future expected returns that the agent receives after following any action in a specific environment state. In continuing tasks with value function approximation (VFA), average rewards and discounted returns can be used as the return formulation but it is unclear how the two formulations compare empirically. This dissertation aims at empirically comparing the two return formulations. We experiment with three continuing tasks of varying complexity, three learning algorithms and four different VFA methods. We conduct three experiments investigating the average performance over multiple hyperparameters, the performance with near-optimal hyperparameters and the hyperparameter sensitivity of each return formulation. Our results show that there is an apparent performance advantage in favour of the average rewards formulation because it is less sensitive to hyperparameters. Once hyperparameters are optimized, the two formulations seem to perform similarly.

Training

Usage

python search.py [options]

For documentation on algorithm parameters refer to Thesis

Option	Description	Default
--num_processes	Number of asynchronous agents	16
--steps	Total number of steps distributed across all synchronous agents in millions	16
--algorithm	Algorithm: Q (for Q-Learning) or SARSA (for SARSA) or doubleQ (for Double Q-Learning)	Q
--network	Network specification: linear or deep (Architecture may depend on task and the degree parameter. See Networks.py for detailed architecture)	linear
--degree	Polynomial Degree to expand environment vector 1, 2 or 3	1
--reward	Type of Reward: discounted for discounted returns or average for average rewards	discounted
--task	The task ID: 1, 2 or 3	1
--lr	Learning Rate	0.0001
--beta	Beta: Used to calculate average reward when reward option is average	0.001
--df	Discount Factor: Used to weight future rewards when reward option is discounted	0.99
--seed	Random seed for experiment replication	0

Logging

The command above will generate logs in the following directory: ./Code/logs/{algorithm}/{reward}/{network}/{process id} The directory will contain:

• A log file for each asynchronous agent containing the reward and total avereage reward at each step
• The network parameters saved for each million steps (cumulated over all agents)
• A hyper_params file specifying:
  • The parameters chosen for the run
  • The Average reward for each saved network (over 5 sample runs of 50,000 steps)

Vizualization

After training the agent, it may be useful to observe its behavior in the environment. To do so use python visualize.py [options]

Option	Description	Default
--task	The task ID: 1, 2 or 3	1
--network	Network specification: linear or deep	linear
--degree	Polynomial Degree to expand environment vector 1, 2 or 3	1
--param	Network parameters path

Note: Vizalusation uses Matplotlib with Qt5Agg backend. Some issues have been identidied on some platforms