Auction-Based Online Policy Adaptation for Evolving Objectives

Dimensions overall score 4.0

GitHub Code Pulse

No public code linked for this paper yet.

• The proposed auction-based mechanism enables dynamic and interpretable trade-offs among multiple objectives in reinforcement learning.method 90%
  Evidencepartial

  The highest bidder selects the action, enabling a dynamic and interpretable trade-off among objectives.

  Implicationpartial

  Directly stated in abstract with clear description of mechanism

  Verificationpartial

  partial

• The system can adapt to changing objectives at runtime by simply adding or removing corresponding policies.method 90%
  Evidencepartial

  when objectives change, the system adapts by simply adding or removing the corresponding policies.

  Implicationpartial

  Explicitly stated in abstract as a core feature of the method

  Verificationpartial

  partial

• Identical copies of a parameterized policy can be deployed for objectives from the same family, facilitating immediate adaptation.technical 85%
  Evidencepartial

  as objectives arise from the same family, identical copies of a parameterized policy can be deployed, facilitating immediate adaptation at runtime.

  Implicationpartial

  Directly stated in abstract with clear technical rationale

  Verificationpartial

  partial

• Selfish local policies are computed by turning the problem into a general-sum game where policies compete against each other.method 90%
  Evidencepartial

  We show how the selfish local policies can be computed by turning the problem into a general-sum game, where the policies compete against each other to fulfill their own objectives.

  Implicationpartial

  Explicitly described in abstract as the computational approach

  Verificationpartial

  partial

• Policies must learn to produce calibrated bids that reflect relative priority to succeed in the auction mechanism.technical 85%
  Evidencepartial

  each policy must not only optimize its own objective, but also reason about the presence of other goals and learn to produce calibrated bids that reflect relative priority.

  Implicationpartial

  Directly stated in abstract as a requirement for success

  Verificationpartial

  partial

• The method achieves substantially better performance than monolithic policies trained with PPO on evaluated tasks.result 85%
  Evidencepartial

  Our method achieves substantially better performance than monolithic policies trained with PPO.

  Implicationpartial

  Directly stated in abstract with comparison to baseline method

  Verificationpartial

  partial

• The framework is evaluated on Atari Assault and a gridworld-based path-planning task with dynamic targets.result 90%
  Evidencepartial

  We evaluate on Atari Assault and a gridworld-based path-planning task with dynamic targets.

  Implicationpartial

  Explicitly stated in abstract as evaluation domains

  Verificationpartial

  partial

• Policies are trained concurrently using proximal policy optimization (PPO) in the implementation.method 90%
  Evidencepartial

  In our implementation, the policies are trained concurrently using proximal policy optimization (PPO).

  Implicationpartial

  Explicitly stated in abstract as the training algorithm used

  Verificationpartial

  partial