ArtEco
A small artificial economy where agents learn to gather, barter and signal without prices or a language given to them.
- Period
- 2026
- Role
- Independent project
- Status
- Experiment
- mean reward when trade is removed
- −92.6%
- mean reward when trade is removed
- message entropy (of a 4-bit maximum)
- 3.97 bits
- message entropy (of a 4-bit maximum)
- information the messages carry about wealth
- ≈ 0.01 bits
- information the messages carry about wealth
Abstract
ArtEco is a 20×20 grid economy in which independent reinforcement-learning agents collect three resources, propose and accept barter with neighbours, and broadcast a 4-bit message, with no prices, market rules or meanings given in advance. Six experiment families (scaling, ablation, language, market equilibrium, supply shock and specialization) were run in pure NumPy, and every number below is recomputed from the published per-episode results. Trade is the load-bearing mechanism: removing it lowers mean episode reward from 319.9 to 23.6 (−92.6%). The message channel is fully used (3.97 of 4 bits of entropy) but carries almost no information about an agent's wealth (mutual information ≈ 0.01 bits), random messages perform as well as learned ones, and across five seeds agents without messages earn more (441.6 ± 103.9 against 309.9 ± 41.9; t = −10.4). Wealth inequality peaks at 50 agents (Gini 0.330) and production is mildly specialized (9.2% below uniform entropy). The main finding is negative: under this training setup an emergent channel appears without becoming a language. Runs are short and single-seed except where stated.
1 Question
Can independent learners discover barter and a shared signal when nothing about prices, markets or meaning is programmed in, and what do the signals actually carry?
2 Environment
A 20×20 grid with three resources (food, metal, energy) at an initial density of 0.15, regenerating with probability 0.03 per cell and step. Each agent observes its position, its inventory, the resources in its cell and every agent's last 4-bit message, and chooses one of eight actions: move in four directions, collect, propose barter to the nearest neighbour, accept a barter, or broadcast. Reward follows survival and resource balance.
Each agent is an independent actor-critic network in plain NumPy: a two-layer 128-unit encoder with an action head (8-way softmax), a message head (4 sigmoids) and a value head, trained with REINFORCE and an advantage baseline (γ = 0.99, learning rate 3·10⁻³). No GPU or framework is needed.
3 Experiments
| Family | Varied | Agents | Episodes |
|---|---|---|---|
| Scaling | N ∈ {10, 30, 50, 100} | 10–100 | 30, 25, 15, 10 |
| Ablation | full / no messages / no trade / random messages | 15 | 35 each |
| Language | entropy, mutual information | 15 | 35 |
| Market | price-ratio convergence | 20 | 30 |
| Shock | metal supply −70% at step 60 | 15 | 30 |
| Specialization | production entropy | 20 | 30 |
4 Results
4.1 Trade carries the economy
- Full model319.9
- No messages416.8
- No trade23.6
- Random messages335.0
Removing trade collapses reward. Removing the message channel does not lower it.
Data
| Condition | Mean reward | Mean trades | Final message entropy |
|---|---|---|---|
| Full model | 319.9 | 72.7 | 3.967 bits |
| No messages | 416.8 | 96 | 0 bits |
| No trade | 23.6 | 0 | 3.966 bits |
| Random messages | 335 | 75 | 3.984 bits |
Without barter, mean reward collapses from 319.9 to 23.6 per episode (−92.6%). Full, random-message and no-message agents trade 72.7, 75.0 and 96.0 times per episode respectively; no-trade agents by construction never do.
4.2 A channel, not a language
Learned messages reach 3.97 of a possible 4 bits of entropy: every symbol is used. But the mutual information between an agent's message and its wealth state is about 0.01 bits, and replacing learned messages with random noise leaves reward unchanged (335.0 against 319.9). Across five seeds, agents without messages earned more:
- With messages309.9
- Without messages441.6
Mean final reward over five seeds (t = −10.4).
Data
| Condition | Mean final reward | Std across 5 seeds |
|---|---|---|
| With messages | 309.9 | 41.9 |
| Without messages | 441.6 | 103.9 |
Communication made outcomes more consistent across seeds (standard deviation 41.9 against 103.9) while lowering the average. High entropy here shows that the channel is used, not that it means anything.
4.3 Scale and inequality
- 10 agents0.238
- 30 agents0.280
- 50 agents0.330
- 100 agents0.254
Data
| Agents | Episodes | Final Gini | Trades in last episode |
|---|---|---|---|
| 10 | 30 | 0.238 | 58 |
| 30 | 25 | 0.28 | 205 |
| 50 | 15 | 0.33 | 421 |
| 100 | 10 | 0.254 | 629 |
Inequality rises from N = 10 to N = 50 and falls again at N = 100, but the larger runs are also shorter (10–15 episodes), so the shape cannot be separated from training length.
4.4 Market and specialization
Market efficiency, realized against possible trades, stays below 1% in every run. Production entropy settles at 1.44 bits against a uniform 1.58 for three resources: agents specialize only slightly (9.2% below uniform). In the shock run (−70% metal at step 60 of each episode) episode rewards show no clear drop, so recovery time cannot be measured from these data.
5 Limits
Every family is a single seed except the five-seed message comparison. Runs are short (10–35 episodes). Mutual information is measured only against wealth state, so the messages may still encode something else; that needs intervention tests that were not run. The reward design, not only learning, shapes how much agents trade.
6 Conclusion
Independent learners in ArtEco discover barter, and the economy depends on it. They also fill a 4-bit channel to near-maximum entropy, yet the channel does not carry wealth information and does not help reward. The honest result is a working barter economy with an unused language.