Asymptotic behaviour of a learning algorithm (Q3319780)
From MaRDI portal
!
This is the item page for this Wikibase entity, intended for internal use and editing purposes. Please use the normal view instead:
scientific article; zbMATH DE number 3850479
| Language | Label | Description | Also known as |
|---|---|---|---|
| default for all languages | No label defined |
||
| English | Asymptotic behaviour of a learning algorithm |
scientific article; zbMATH DE number 3850479 |
Statements
Asymptotic behaviour of a learning algorithm (English)
0 references
1984
0 references
learning algorithm
0 references
stochastic automata
0 references
learning system
0 references
linear reward- epsilon penalty algorithm
0 references
asymptotic properties
0 references
stationary random environment
0 references
action probability
0 references
Gauss-Markov diffusion process
0 references
0 references
0.8334971070289612
0 references
0.815601646900177
0 references
0.8117983937263489
0 references
0.8003101944923401
0 references
Weak convergence ideas are used in proving asymptotic properties of a stochastic automaton with \(r(\geq 2)\) actions, acting in unknown stationary random environment. New techniques for analyzing the linear reward-\(\epsilon\) penalty algorithm for the multiple action case are applied and the analysis of the asymptotic properties of the action probability associated with the best action is made. The set up is Markovian and the tightness of the resulting sequence of random variables for small values of parameters in the updating algorithm is investigated. The evolution of the system can be asymptotically described by a Gauss- Markov diffusion process whose parameters can be explicity computed. The diffusion so obtained is also seen to be stationary. It appears that similar results can be obtained if a linear reward-inaction algorithm is used, but the action probabilities are not permitted to go to the extreme values of zero and unity.
0 references