# MCTS guided Genetic Algorithm for optimization of neural network weights

**Authors:** Akshay Hebbar  
**arXiv:** [2308.04459](https://arxiv.org/abs/2308.04459)  
**Date:** 2023-08-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.04459v1/

## Key points

- Research investigates applying a search strategy to genetic algorithms for optimizing neural network weights
- Genetic algorithms commonly used for various problem domains, but focus here is on neural network weights optimization
- Simpler tree search algorithms (breadth-first, depth-first, iterative) computationally heavy and result in long execution times
- Proposed solution: combine genetic algorithms with Monte Carlo Tree Search (MCTS)
- MCTS originally developed for game-tree searches, calculates cumulated rewards from child nodes using Q and N values
- Exploration approach considers number of visited nodes to discover unvisited child nodes
- Expansion approach follows qualitative strategy to discover child nodes with high Q-values indicating cumulative rewards
- UCT (upper confidence trees) found to be sufficient policy for MCTS based on prior research, balances exploration versus exploitation
- Experimental results presented using heatmap of diabetes dataset with different mutation rates and crossover probabilities applied to genetic algorithms
- Combining genetic algorithms with MCTS can optimize neural network weights more efficiently compared to traditional tree search methods
- Approach holds promise for improving optimization processes in various problem domains involving neural networks

## Summary

In this research, the authors investigate the possibility of applying a search strategy to genetic algorithms in order to optimize neural network weights. Genetic algorithms (GA) are commonly used for various problem domains, but in this paper, the focus is specifically on optimizing neural network weights. The authors highlight that simpler tree search algorithms such as breadth-first, depth-first and iterative techniques can be computationally heavy and result in long execution times. To address this issue, they propose combining genetic algorithms with Monte Carlo Tree Search (MCTS), which has proven to be an effective tree search strategy given states and rewards. The authors provide an overview of the MCTS approach, which was originally developed for game-tree searches. MCTS works by calculating cumulated rewards from child nodes and using Q and N values to balance between exploration and exploitation approaches. The exploration approach considers the number of visited nodes and uses a quantitative approach to discover child nodes that have not been visited. On the other hand, the expansion approach follows a qualitative strategy to discover child nodes with high Q-values indicating cumulative rewards. The authors mention that UCT (upper confidence trees) has been found to be a sufficient policy for MCTS based on prior research. UCT provides an upper confidence bound to tree search and helps balance exploration versus exploitation. To demonstrate their proposed approach, the authors present experimental results in optimizing neural network weights using genetic algorithms guided by MCTS. They show a heatmap of the diabetes dataset where different combinations of mutation rates and crossover probabilities are applied to genetic algorithms. Overall, this research explores how combining genetic algorithms with MCTS can optimize neural network weights more efficiently compared to traditional tree search methods. The experimental results suggest that this approach holds promise for improving optimization processes in various problem domains involving neural networks.