Hi all, I have the following questions:
1) Does the solution with meta-heuristic (say genetic algorithm or similar) or reinforcement learning converges faster for the same optimization problem?
2) Is the solution obtained using reinforcement learning an optimal one?
Thanks in advance
Dear Samundra,
There are improvements in performance when hybridizing these two types of algorithms in both directions, in order to verify this I suggest you to see links and attached files on topic.
Reinforcement learning for solution updating in Artificial Bee Colony
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6049945/
The International Conference on Advanced Machine Learning ...
https://books.google.dz/books?isbn=3030141187
Survey of Meta-Heuristic Algorithms for Deep Learning Training ...
https://www.intechopen.com/.../optimization-algorithms.../survey-o...
Deep Reinforcement Learning Demysitifed (Episode 2) — Policy ...
https://medium.com/.../deep-reinforcement-learning-demysitifed-e...
Learning Combinatorial Optimization Algorithms ... - NIPS Proceedings
https://papers.nips.cc/.../7214-learning-combinatorial-optimization-...
Best regards
The answer really depends on how efficient the models are working, as reflected by the coding of the search process.
It depends on the way you are going to compare, the type of the application, the scale of the search space and the way you balance exploitation vs exploration of the search space. Naturally, meta-heuristic optimization tries to find a feasible solution not the optimal (deterministically the best) one. For ultra-large search spaces there are parameters to control the convergence speed, if you force the algorithm to converge too fast, it might get caught in local optima. Multi-agent meta-heuristic optimization strategies (ACO, Bee colony, etc.) work well for large search spaces or dynamic environments if the solution and the environment are both appropriately modeled.
Hi Samundra,
I would say that Genetic Algorithms (GA) and the like are used for different problems than Reinforcement Learning (RL). A GA generally is used for finding an optimum (point in n-dimensional space) of a (complicated) goal function. RL on the other hand is used to determine an optimal policy (recipe for acting) in an environment, such a playing a game or interacting with other 'agents'. So it really depends on the type of problem you have.
Both techniques do not garantee finding a (global) optimal solution, except maybe under very ideal circumstances.
- Badges
- Science topic