search

Niketkumardheeryan / ML-CaPsule

ML-capsule is a Project for beginners and experienced data science Enthusiasts who don't have a mentor or guidance and wish to learn Machine learning. Using our repo they can learn ML, DL, and many related technologies with different real-world projects and become Interview ready.
MIT License
418 stars 356 forks source link

Waste Management through Reinforcement Learning #1142

Open Panchadip-128 opened 1 month ago

Panchadip-128 commented 1 month ago

The project aims to develop a reinforcement learning (RL) agent to optimize waste collection in a simulated environment, minimizing overflow events and improving efficiency.

Environment and State Representation: The state is represented by four features: Waste Level: Current waste level (0 to 1) Time of Day: A random value representing the time (0 to 24 hours) Weather Condition: A random value (0 to 1) indicating the weather Distance to Collection Point: A random value (0 to 10) representing the distance to the waste collection point.

Action Space: The agent can choose between two actions: Wait (0): Do not collect waste. Collect Waste (1): Proceed with waste collection.

Reward Structure: The reward system is designed to encourage efficient waste collection: +10 for timely collection when the waste level exceeds the threshold. -5 for premature collection when the waste level is below the threshold. -1 for each time step to penalize waiting.

Training Process: The agent is trained over 100 episodes, where each episode simulates a series of time steps (up to 20) where the agent makes decisions based on the current state. The agent learns from experience using a replay memory and updates its policy through Q-learning.

Evaluation Metrics: Performance is evaluated using: Average Reward per Episode: Measures the effectiveness of the agent's actions. Epsilon Decay: Tracks the exploration rate, indicating how the agent balances exploration vs. exploitation. Overflow Events: Counts occurrences when the waste level exceeds the maximum capacity as per previous updation.

Visualization: The results are visualized using Matplotlib to plot: Average rewards per episode, showing the agent's learning progression and rewards gained on successfull execution and implementation of a specified condition Epsilon decay over episodes, illustrating the shift from exploration to exploitation. Overflow events per episode, highlighting improvements in waste management techniques

wm3 wm2 wm1

Panchadip-128 commented 1 month ago

@Niketkumardheeryan Please review this once

Niketkumardheeryan commented 4 weeks ago

go for it