Bridging Two Realms of Machine Ethics

Authors - Luís Moniz Pereira and Ari Saptawijaya (2016)

​

Presentation By - Manish Milind Adkar

What is Machine Ethics?

• Adds/ensures moral behavior in artificial intelligent agents.
• Focuses on
• The AI control problem
• Bias in machine learning
• Autonomous systems
• Integration of Artificial General Intelligence in society
• Ultimate goal is to create a machine that itself will follow an ideal ethical set of principles and make decisions accordingly.
• Also, helps us better understand human morality.

Introduction

• Various techniques to address problems in machine ethics.
• Machine learning (case based reasoning, artificial neural networks, etc.)
• Logic-based formalisms (deontic logic, non-monotonic logic, etc.)
• Research has a focus on logic programming techniques and their appropriateness to model the morality aspects.
• Mostly deals with moral permissibility and its justification.

Brief about the Two Realms

• Individual Realm:
• Talks about modeling the behavior and moral cognition of an individual agent.
• Logic programming based agent architecture.

​

• Collective Realm:
• Talks about the simple minded agents (without cognitive capabilities) in a population.
• Evolutionary Game Theory.

​

• Bridging the Two Realms:
• Modeling moral cognition in individuals of a population and its effects on their strategies.
• Help in understanding the emergent behavior of ethical agents in groups.

The Individual Realm

• Computational Approaches
• Jeremy (advisor system) follows utilitarianism. Decisions based on total net pleasure.
• Machine learning:
• MedEthEx and EthEl (advisor systems) related to medicine.
• Case based reasoning in TruthTeller and SIROCCO to give info related to ethical dilemmas.
• Artificial Neural Networks to understand morality from the philosophy of ethics viewpoint.
• Deontic logic.
• Category theory used as meta-level of moral reasoning.

The Individual Realm

• Logic Programming
• LP features like abduction, updating, preferences, etc. make it very suitable for machine ethics.
• Abduction is a reasoning method to infer the best-preferred explanation to observed evidence.
• Addressed the issue of moral permissibility using Doctrine of Double Effect (DDE) as the basis. (distinguishing whether a harm consequence is merely a side-effect of achieving a good result or rather a means to bring about the good result.
• Further extended work considers the Doctrine of Triple Effect (DTE) which distinguishes further whether doing an action in order that the harm occurs and doing it because that harm will occur.
• Trolley problem (and its various cases).

Trolley Problem

1. Classic
2. Loop case
3. Loop-Push case

• The integrated LP approach successfully delivers the moral decisions for these scenarios which also conforms to the experimental study by Hauser.
• Different outcomes between DDE and DTE.
• Loop case:
• Diverting trolley morally impermissible in DDE while permissible in DTE.
• Permissible by DTE as the trolley will hit the man but not in order to intentionally kill him.
• Loop-Push case:
• Pushing the man is morally impermissible in both DDE and DTE.
• Impermissible as the deliberate action of pushing is considered performed in order to bring harm.

• Appropriateness of LP is justified using counterfactuals.
• Counterfactuals: process of reasoning about a past event that did not occur to reason about a past event that did occur. E.g. “If I hadn’t taken a sip of this hot coffee, I wouldn’t have burned my tongue”.
• If E (a harm) would not have been true, then G (a goal) would not have been true. If the counterfactual is valid, then E is instrumental as a cause of G, and not a mere side-effect of the action. Since E is morally wrong, achieving G that way, by means of that action, is impermissible; otherwise, it is not.
• Three step process:
• Abduction to explain the current observation and fixes context in which the counterfactual is evaluated by means of LP updating.
• Causal intervention is realized by hypothetical updates.
• The modified program is now verified to see if the counterfactual holds true.

• Loop case:
• For DDE, the counterfactual “if the man had not been hit by the trolley, the five people would not have been saved” is evaluated. As harm to the man is instrumental for the goal of saving five, hence diverting the trolley is impermissible.
• For DTE, two counterfactuals are evaluated. First, “if the man had not been on the side track, then he would not have been hit by the trolley” is verified whether the man getting hit was a consequence of him being on the that sidetrack. Second, “if the man had not been pushed, then he would not have been hit by the trolley” (assumed to place the man on the sidetrack) but is not valid, hence diverting the trolley is permissible.
• Loop-Push case:
• For DDE, the previous counterfactual is still valid and hence not permissible.
• For DTE, now the second counterfactual is valid, and hence the action of pushing and diverting is now impermissible.

Individual Realm

• LP updating is appropriate for representing changes and for dealing with incomplete information. It can be used for the adoption of new ethical rules on top of the existing ones the agent currently follows.
• ACORDA and Evolution Prospection Agent (EPA) are examples of implemented systems which benefits from the LP features.
• Hence, individual agents can be endowed with the capability to declaratively represent ethical situations so they can reason on ethical issues arising from such situations.

The Collective Realm

• The mechanisms of emergence and evolution of cooperation in populations of abstract individuals is studied via the Evolutionary Game Theory (EGT).
• In such collective strategic interaction environments, conflicts arise due to actions of individual agents on the welfare of others (also their own in return).
• The research describes the spontaneous emergence of order approach where norms are established from endogenous agreements.
• In simple terms, a cooperative act is the act of paying cost to convey a benefit to someone else.
• The study is useful in establishing cooperative capabilities for spontaneous organizations and control of swarm of autonomous robots.

Intention Recognition

• Promotes the evolution of cooperation and hence plays an important role in many cooperative interactions.
• Recognized intents can be used to understand how another player will play the next round (cooperate / defect).
• A study of a population of intention recognizers was conducted, implemented as a Bayesian Network, used info from data gathered from previous one-on-one experiences and current intent signals.
• Results suggest, incorporating such agents in the population improves the cooperation and reduces misunderstandings.

Commitments

• Promotes the emergence of cooperation.
• Agents ask for commitments to others and provide incentives in order to ensure a combined good outcome.
• A study evaluates whether costly commitment strategies are viable for cooperation using the Prisoner’s Dilemma game.

Prisoner’s Dilemma

Here, “stays silent” is cooperate (C)�and “betrays” is defect (D).

​

A can propose B prior to the game to play

C and is willing to pay a personal cost.

• Free-riding strategies added in the model.
• Fake committers.
• Commitment free riders.
• Fake committers can be handled with strong compensation cost.
• Commitment free-riders will keep defecting among themselves while the commitment proposers will cooperate with similar players.

​

• Intention recognition can also be used in collusion to avoid the high commitment costs when it was useless.

Apology in committed vs. commitment-free repeated interactions

• Most powerful and ubiquitous method of conflict resolution.
• Implemented in several HCI systems.
• Iterated Prisoner’s Dilemma is a standard model to investigate conflict resolution.

Iterated Prisoner’s Dilemma

• An experiment was conducted with a model containing strategies that explicitly apologizes (with high cost) for errors between rounds.
• A population consisting of only apologizers can maintain perfect cooperation.
• Fake apologizers could emerge, who would just accept apologies but won’t apologize for their mistakes.
• Hence, commitments are also necessary before engaging in interactions.
• Apologies itself cannot guarantee evolution of cooperation.

​

• One of the studies also provide insights on design and deployment of such mechanisms in real world. Like what kind of apologies to make in which situation and whether apologies/compensation can be enhanced for better experience.

Bridging the Realms

• We saw that individual realm deals with cognition and reasoning in individual agents to compete amongst free riders and deceivers.
• The collective realm deals with strategies to enhance cooperation and setup norms through evolution amongst a group of individuals.
• Can the results from the study of individual cognition be equally applicable to the evolution of population of such agents (and vice versa)? - Yes
• Development of cooperation induces the emergence of morals and not the other way around.

References

• Paper: https://userweb.fct.unl.pt/~lmp/publications/online-papers/Bridging%20Two%20Realms%20of%20Machine%20Ethics.pdf
• Iterated Prisoner’s Dilemma: https://www.youtube.com/watch?v=BOvAbjfJ0x0
• Trolley Problem: https://www.youtube.com/watch?v=yg16u_bzjPE

​

Thank you