kth.sePublications KTH
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Frameskipping and Exploration Strategies for Deep Q-Networks
KTH, School of Electrical Engineering and Computer Science (EECS).
KTH, School of Electrical Engineering and Computer Science (EECS).
2024 (English)Independent thesis Basic level (degree of Bachelor), 10 credits / 15 HE creditsStudent thesis
Abstract [en]

This paper studies the impact of frameskipping strategies and explorationpolices on the efficiency and the performance of Deep Q-Network (DQN) agents in theArcade Learning Environment (ALE), using the game Space Invaders. Two frameskippingstrategies are examined: fixed steps and fixed episodes. In fixed steps training, agents aretrained for a set number of steps, resulting in a fixed number of experiences. In fixed episodestraining, agents are trained until a set number of episodes is completed, resulting in varyingnumbers of experiences. The findings suggest that higher frameskips shorten DQN trainingwithout performance loss when utilizing the strategy to train over fixed episodes. Conversely,higher performance is achieved by increasing frameskip when training over fixed steps. Thissuggests that the amount of episodes experienced by the DQN is of significant importance toincrease performance. In addition, two agents trained on different exploration policies ε-greedy exploration policy and Boltzmann exploration policy, are compared. The results showthat ε-greedy implementation outperformed Boltzmann exploration, which might be onaccount of the sensitivity of the Boltzmann temperature parameter τ .

Abstract [sv]

Den här rapporten undersöker påverkan av frameskipping strategieroch utforskning policyer har på effektiviteten och prestandan av agenter tränade med Djup Q-inlärning (DQN) för Arcade Learning Environment (ALE), i spelet Space Invaders. Tvåframeskipping strategier undersökes: fasta steg och fasta episoder. Fasta steg innebär attagenten tränas ett fast antal steg, vilket resulterar i en konstant mängd erfarenheter. Fastaepisoder innebär att agenten tränar ett fast antal episoder med varierande numer averfarenheter. Resultaten tyder på att högre frameskip kan användas för att efektiviseratränigen för en DQN utan att kompromissa prestandan genom att använda strategin över fastaepisoder. Och omvänt så kan högre prestanda fås genom träning över fasta steg. Detta tyderpå att mängden episoder som DQN uplever är av stor betydelse för prestandan. Dessutomimplementeras ε-girig utfrorsknings policy och Boltzmann utforsknings policy i syfte att göraen generell jämförelse mellan de två. Resultaten återspeglar att ε-girig utforsknings policypresterarade bättre än Boltzmanns utforsknings policy, antagligen på grund av känsligheten avBoltzmann temperaturparametern τ.

Place, publisher, year, edition, pages
2024. , p. 65-74
Series
TRITA-EECS-EX ; 2024:136
Keywords [en]
Frameskipping, Deep Q-Networks(DQN), Arcade Learning Environment (ALE), Space Invaders, Boltzmann, ϵ-Greedy
National Category
Electrical Engineering, Electronic Engineering, Information Engineering
Identifiers
URN: urn:nbn:se:kth:diva-357842OAI: oai:DiVA.org:kth-357842DiVA, id: diva2:1922200
Supervisors
Examiners
Projects
Kandidatexamensarbete i Elektroteknik 2024Available from: 2024-12-18 Created: 2024-12-18

Open Access in DiVA

fulltext(121150 kB)95 downloads
File information
File name FULLTEXT01.pdfFile size 121150 kBChecksum SHA-512
82342f1408fe2aae929f55e76f2a176a8521cf94c0fe100464225724a9b74ddff6f61a0fb1cf957b5d6400be30877b9d0aec28080253ae7c1e09793536e10217
Type fulltextMimetype application/pdf

By organisation
School of Electrical Engineering and Computer Science (EECS)
Electrical Engineering, Electronic Engineering, Information Engineering

Search outside of DiVA

GoogleGoogle Scholar
Total: 95 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 331 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf