Chronocooked: A Benchmark for Implicit Interval Tim... | AI Research

Chronocooked is a reinforcement learning (RL) benchmark suite designed to evaluate how artificial agents develop an internal sense of time. While many AI systems perform tasks involving temporal sequences, they often rely on external cues or task-specific heuristics rather than an implicit understanding of elapsed time. This benchmark introduces cooking scenarios where temporal information is unobserved, requiring agents to infer time to achieve optimal performance.

The Chronocooked Environment

The environment is a simplified, single-agent grid world inspired by the Overcooked game. An agent must pick up an onion, place it in an oven, and wait for it to cook. The oven’s timer is invisible to the agent, meaning the agent must learn to estimate the duration required for the soup to be ready. The researchers designed this environment to be simple enough to support controlled experiments with various model architectures, including non-recurrent, recurrent, and biologically plausible models.

Experimental Tasks

The benchmark includes several tasks adapted from psychology to test different aspects of temporal decision-making:

• Bisection Task: Agents must categorize intervals as "short" or "long." This tests how agents encode time and whether they replicate human-like biases, such as the tendency to perceive time logarithmically or linearly.

• Fixed Interval Timing: Agents are rewarded for retrieving soup only after a specific target duration has passed. This measures the agent’s ability to control its actions based on internal time-keeping.

• Multi-Timer and Uncertainty Tasks: These tasks evaluate an agent’s capacity to manage multiple target durations simultaneously, generalize to durations not seen during training, and adapt to scenarios where the oven might not transition to a "ready" state automatically.

Performance and Model Comparison

The researchers evaluated three model architectures: a non-recurrent model (CNN-MLP), a recurrent model (CNN-LSTM-MLP), and a biologically plausible recurrent model (CNN-CTRNN-MLP).
Key findings include:

• Emergent Timing: Non-recurrent models typically checked the oven immediately and repeatedly, showing no internal sense of time. Recurrent models, however, exhibited emergent time-keeping behavior.

• Model Differences: The LSTM-based models generally performed better than the biologically plausible CTRNN models at lower memory sizes, though performance gaps narrowed as memory capacity increased.

• Limitations in Human-like Timing: While the models could discriminate between intervals, they struggled to fully replicate human timing characteristics. Specifically, while the mean response times increased linearly with target durations, the models failed to conform to the "scalar property" of variance, where the ratio of variance to mean remains constant.

Why This Matters

The authors argue that incorporating time perception is essential for artificial agents intended for human-robot interaction. In real-world societies, humans operate based on implicit temporal expectations. By providing a standardized benchmark, the researchers aim to move beyond simple task performance and encourage the development of agents that can navigate time-dependent environments with a more robust, internal sense of duration.