Synthesising Reinforcement Learning Policies Through Set-Valued Inductive Rule Learning

Coppens, Youri; Steckelmacher, Denis; Jonker, Catholijn M.; Nowé, Ann

doi:10.1007/978-3-030-73959-1_15

Cited by 2 publications

(6 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Most of the experiments described in this section focus on quantitatively evaluating the playing strength of the different proposed types of decision trees in a variety of games. This is consistent with most related work on explainable RL, in which researchers tend to take the inherent explainability of models such as decision trees for granted, and empirically compare the performance of such models to the full (less explainable) original policies (Liu et al, 2019;Coppens et al, 2019Coppens et al, , 2021Deproost, 2021).…”

Section: Methodssupporting

confidence: 80%

“…These could, for example, be included in automatically-generated instruction manuals for humans (Stephenson et al, 2022). Coppens et al (2019Coppens et al ( , 2021; Deproost (2021) distil trained policies into various forms of decision trees and rules, which can lead to local (state-specific) as well as global (game-wide) explanations of policies. These were evaluated in environments such as the Mario AI benchmark, Ms Pacman, and Enduro.…”

Section: Related Work On Explainability In Reinforce-ment Learning An...mentioning

confidence: 99%

“…Resulting vectors of logits are transformed into discrete probability distributions over the actions by applying a softmax (plus invalid action masking (Huang and Ontañón, 2020) in games where some actions are sometimes illegal). When such policies are subsequently distilled into decision trees or rules for explainability, these are typically 8.3 DECISION TREES OF STATE-ACTION FEATURES 8 183 trained as classifiers that again use representations of states s as input (splitting on features of states) (Coppens et al, 2019(Coppens et al, , 2021Deproost, 2021). These models then produce probability distributions over all actions (which must all be explicitly enumerated as potential target classes) as outputs.…”

Section: Decision Tree Structuresmentioning

confidence: 99%

“…In related work on explaining policies trained with reinforcement learning (Coppens et al, 2019(Coppens et al, , 2021Deproost, 2021), it is typically assumed that decision trees or rules are inherently interpretable. Empirical evaluations then consist of estimating the change (typically reduction) in a policy's performance when distilling it from a more complex model into a smaller decision tree or set of rules.…”

Section: Extracting Insights From Features and Policiesmentioning

confidence: 99%

“…The new types of decision trees proposed in Chapter 8, which operate on state-action pairs as inputs, are a first step towards extending previous work on explaining RL-based policies (Coppens et al, 2019(Coppens et al, , 2021Deproost, 2021) to domains with large action spaces. Explainability in games (Silva et al, 2021;Pálsson and Björnsson, 2022;Stephenson et al, 2022), search Kaisers, 2020, 2021), RL (Liu et al, 2019;Hilton et al, 2020;Lin et al, 2021), and AI more gen-IMPACT PARAGRAPH 275 erally (Molnar, 2020;Zhang et al, 2021), are increasingly recognised as crucial fields of research.…”

Section: Indexmentioning

confidence: 99%

See 4 more Smart Citations

Learning state-action features for general game playing

Soemers¹

View full text Add to dashboard Cite

published version features the final layout of the paper including the volume, issue and page numbers. Link to publication General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal.If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the "Taverne" license above, please follow below link for the End User

show abstract

Section: Methodssupporting

confidence: 80%

Section: Related Work On Explainability In Reinforce-ment Learning An...mentioning

confidence: 99%

Section: Decision Tree Structuresmentioning

confidence: 99%

Section: Extracting Insights From Features and Policiesmentioning

confidence: 99%

Section: Indexmentioning

confidence: 99%

See 3 more Smart Citations

Learning state-action features for general game playing

Soemers¹

View full text Add to dashboard Cite

show abstract

Inverse reinforcement learning through logic constraint inference

2023

View full text Add to dashboard Cite

Autonomous robots start to be integrated in human environments where explicit and implicit social norms guide the behavior of all agents. To assure safety and predictability, these artificial agents should act in accordance with the applicable social norms. However, it is not straightforward to define these rules and incorporate them in an agent's policy. Particularly because social norms are often implicit and environment specific. In this paper, we propose a novel iterative approach to extract a set of rules from observed human trajectories. This hybrid method combines the strengths of inverse reinforcement learning and inductive logic programming. We experimentally show how our method successfully induces a compact logic program which represents the behavioral constraints applicable in a Tower of Hanoi and a traffic simulator environment. The induced program is adopted as prior knowledge by a model-free reinforcement learning agent to speed up training and prevent any social norm violation during exploration and deployment. Moreover, expressing norms as a logic program provides improved interpretability, which is an important pillar in the design of safe artificial agents, as well as transferability to similar environments.

show abstract

Synthesising Reinforcement Learning Policies Through Set-Valued Inductive Rule Learning

Cited by 2 publications

References 17 publications

Learning state-action features for general game playing

Learning state-action features for general game playing

Inverse reinforcement learning through logic constraint inference

Contact Info

Product

Resources

About