Autonomous Vehicle Decision-Making: Should We Be Bio-inspired?

摘要

On our crowded roads, drivers must compete for space but cooperate to avoid occupying the same space at the same time. Decision-making is strategic and requires mutual understanding of other's choices. Fully autonomous vehicles (AVs) will need risk management software to make these types strategic decisions without human arbitration. Accidents will occur, and what constitutes rational and 'safe' decisions will be scrutinized by the legal system. It is far from clear how AV-Human and AV-AV interactions should be managed. Game Theory provides a framework for analyzing mutual 'games' with 2 or more players. It assumes that players mutually optimize their outcomes according to Nash equilibria (NE), but do humans follow Nash equilibria in Human-Human interactions? We implemented simple two-player competitive games to see whether people played rationally according to Nash equilibria. On each of 100 trials, each player was instructed to maximise their reward by pressing one of three buttons labelled "4", "6", and "12", without knowing the other players choice. If players pressed different buttons, they received a reward of 4, 6, or 12 points accordingly. If players pressed the same button, the reward was reduced depending on the game type. Results showed that players did not follow NE, but played a probabilistic game that included the "4" button, even though pressing this button is always suboptimal. We suggest that this may be an evolutionary strategy, but it clearly shows that people do not follow the 'rational' Nash strategy. It seems that AV-human interactions will be probabilistic. In AV-AV interactions, software may be playing itself, and may also require probabilistic optimal evolutionary-type strategies. We doubt that the full implications of autonomous decision-making have been fully worked out. Whether probabilistic decisions will tolerated legally and actuarially is doubtful. One way to avoid them would be to allow regulated AV-AV communications, and force software decisions to be deterministic according to some protocol. However, AV-Human interactions seem likely to remain problematic.