D'BEAST: Logged Bad Experiences as Explanation for Autonomous Driving Agents
Organizations: Information Technology, Universitas Muhammadiyah Yogyakarta, Yogyakarta, Indonesia. · Division of Information Science, Nara Institute of Science and Technology, Japan. · Department of Industrial and Systems Engineering, IRC Smart Mobility and Logistics, King Fahd University of Petroleum and Minerals (KFUPM), Saudi Arabia.
Abstract
Reinforcement learning agents learn from mistakes, yet their decisions can be difficult to trace back to those experiences. We investigate whether a tabular agent's decisions can be explained by its bad experiences, defined as the training transitions associated with safety or task events. A fixed training log is replayed from the same initial Q-table under three conditions. FULL retains all updates, SKIP omits bad-experience updates, and NEUTRALIZE retains those updates without their event penalties. We illustrate the method on six classic Markov decision processes and evaluate it in closed-loop driving in CARLA. On the full driving task, FULL reached the goal from 33 of 40 test starts, whereas SKIP and NEUTRALIZE reached it from none, indicating that the bad experiences carry decision-relevant information. At the local level, for 47 decisions shortly before a failure of SKIP or NEUTRALIZE, removing or neutralizing a small group of bad experiences, with a median size of one, caused the trained agent to select the failing policy's action in 46 cases. For this agent, the results indicate that its decisions can be explained by its bad experiences.
Figures & tables
| Event | Definition | Penalty | Terminal |
|---|---|---|---|
| Collision | Contact with a vehicle or pedestrian | 100 | yes |
| Road departure | Vehicle centre off the road | 100 | yes |
| Failure to yield | 1.5 m/s or faster within 4 m of a pedestrian on or entering the path | 1 | no |
| Offside excursion | A body corner in the oncoming lane | 1 | no |
| Lane-edge contact | A body corner off the road while the centre is on it | 1 | no |
| Insufficient clearance | An object closer than 2 m ahead or 1 m beside | 1 | no |
| Gate | Test starts (of 40) | |||||
| Stage | Ep. | Bad / all updates | (of 20) | Full | Skip | Neutralize |
| Straight | 150 | 2,742 / 14,461 | 20 | 40 | 40 | 40 |
| Turn | 900 | 20,001 / 78,448 | 18 | 38 | 40 | 40 |
| Loop | 400 | 11,809 / 60,280 | 20 | 40 | 27 | 9 |
| Parked vehicle | 800 | 24,227 / 75,980 | 17 | 38 | 39 | 38 |
| Pedestrian | 600 | 7,346 / 53,667 | 19 | 34 | 32 | 10 |
Interpretability in RL
Interpretability in reinforcement learning makes learned policies, decisions, and internal representations understandable and auditable.
See all 81 papers